diff --git a/.github/workflows/cpp-tests-llm.yml b/.github/workflows/cpp-tests-llm.yml
index b798baadfe..7e2eb69f59 100644
--- a/.github/workflows/cpp-tests-llm.yml
+++ b/.github/workflows/cpp-tests-llm.yml
@@ -37,7 +37,7 @@ jobs:
             platform: linux
             runner: qvac-ubuntu2404-x64
             arch: x64
-          - os: macos-14
+          - os: macos-15-xlarge
             platform: darwin
             arch: arm64
           - os: windows-2025
diff --git a/packages/llm-llamacpp/CMakeLists.txt b/packages/llm-llamacpp/CMakeLists.txt
index ce1601da75..43f84e529d 100644
--- a/packages/llm-llamacpp/CMakeLists.txt
+++ b/packages/llm-llamacpp/CMakeLists.txt
@@ -172,7 +172,7 @@ if(BUILD_CLI)
   target_include_directories(cli_tool PRIVATE ${QVAC_LIB_INFERENCE_ADDON_CPP_INCLUDE_DIRS})
   # Set C++20 standard for std::views support
   target_compile_features(cli_tool PRIVATE cxx_std_20)
-  
+
   # Install ggml backends next to the CLI tool binary in build folder
   if((ANDROID OR UNIX) AND NOT APPLE)
     foreach(_backend ${GGML_AVAILABLE_BACKENDS})
@@ -190,7 +190,7 @@ if(BUILD_TESTING)
   find_package(GTest CONFIG REQUIRED)
   include(GoogleTest)
   enable_testing()
-  
+
   # Integration tests for model classes (includes backend selection tests)
   # Pass ENABLE_COVERAGE option to test subdirectory
   add_subdirectory(test/unit)
diff --git a/packages/llm-llamacpp/addon/src/addon/AddonJs.hpp b/packages/llm-llamacpp/addon/src/addon/AddonJs.hpp
index 0cfe0b7998..512715034c 100644
--- a/packages/llm-llamacpp/addon/src/addon/AddonJs.hpp
+++ b/packages/llm-llamacpp/addon/src/addon/AddonJs.hpp
@@ -1,6 +1,8 @@
 #pragma once
 #include <any>
+#include <cmath>
 #include <functional>
+#include <limits>
 #include <memory>
 #include <string>
 #include <type_traits>
@@ -371,15 +373,18 @@ inline void parseGenerationParams(
   auto reasoningBudget = configObj->getOptionalPropertyAs<js::Number, double>(
       env, "reasoning_budget");
   if (reasoningBudget.has_value()) {
-    // Exact compare (0 and -1 are representable doubles) so fractional
-    // inputs like 0.5 are rejected, not silently truncated.
-    if (*reasoningBudget != 0 && *reasoningBudget != -1) {
+    // Reject fractional inputs (0.5, -1.1, 32.7) by requiring the value to
+    // round-trip through int. -1 = unrestricted, 0 = disabled, N>0 caps the
+    // reasoning channel at N tokens via the budget sampler.
+    const double value = *reasoningBudget;
+    if (value < -1 || value != std::floor(value) ||
+        value > static_cast<double>(std::numeric_limits<int>::max())) {
       throw StatusError(
           general_error::InvalidArgument,
-          "generationParams.reasoning_budget must be -1 (unrestricted) "
-          "or 0 (disabled)");
+          "generationParams.reasoning_budget must be -1 (unrestricted), "
+          "0 (disabled), or a positive integer (token cap)");
     }
-    overrides.reasoning_budget = static_cast<int>(*reasoningBudget);
+    overrides.reasoning_budget = static_cast<int>(value);
   }
 }
 
diff --git a/packages/llm-llamacpp/addon/src/model-interface/GenerationParamsApply.cpp b/packages/llm-llamacpp/addon/src/model-interface/GenerationParamsApply.cpp
index 815615b5da..fd8271e7b7 100644
--- a/packages/llm-llamacpp/addon/src/model-interface/GenerationParamsApply.cpp
+++ b/packages/llm-llamacpp/addon/src/model-interface/GenerationParamsApply.cpp
@@ -29,6 +29,21 @@ void applyGenerationOverridesToSampling(
   setIf(overrides.presence_penalty, sampling.penalty_present);
   setIf(overrides.repeat_penalty, sampling.penalty_repeat);
 
+  // Forward reasoning_budget into the budget-sampler's token cap. The
+  // template-specific start/end/forced vectors are refreshed after prompt
+  // formatting, where common_chat_params exposes the thinking tags.
+  if (overrides.reasoning_budget) {
+    const int budget = *overrides.reasoning_budget;
+    if (budget > 0) {
+      sampling.reasoning_budget_tokens = budget;
+    } else {
+      sampling.reasoning_budget_tokens = -1;
+      sampling.reasoning_budget_start.clear();
+      sampling.reasoning_budget_end.clear();
+      sampling.reasoning_budget_forced.clear();
+    }
+  }
+
   // `json_schema` and `grammar` are mutually exclusive at the JS boundary
   // and in `AddonJs::runJob::parseText`, so reaching this branch with both
   // set means a caller bypassed those checks (most likely the C++ unit
diff --git a/packages/llm-llamacpp/addon/src/model-interface/LlamaModel.cpp b/packages/llm-llamacpp/addon/src/model-interface/LlamaModel.cpp
index fe1b8c500a..3253145cc7 100644
--- a/packages/llm-llamacpp/addon/src/model-interface/LlamaModel.cpp
+++ b/packages/llm-llamacpp/addon/src/model-interface/LlamaModel.cpp
@@ -98,7 +98,8 @@ void LlamaModel::resolveShardPaths(
 void LlamaModel::tuneConfigMap(
     std::unordered_map<std::string, std::string>& configFilemap,
     const ModelMetaData& metadata, const std::optional<int>& adrenoVersion,
-    const FinetuneConfigOverrides& finetuneOverrides, bool isOpenCl) {
+    const FinetuneConfigOverrides& finetuneOverrides, bool isOpenCl,
+    bool isMetal) {
 
   const bool isFinetuning = finetuneOverrides.active;
 
@@ -143,13 +144,11 @@ void LlamaModel::tuneConfigMap(
     QLOG_IF(
         Priority::INFO,
         "[LlamaModel] BitNet model detected: disabling flash attention\n");
-  } else if (isOpenCl && notUserSet("flash-attn", "flash_attn")) {
+  } else if (notUserSet("flash-attn", "flash_attn")) {
     configFilemap.erase("flash_attn");
-    configFilemap["flash-attn"] = "off";
+    configFilemap["flash-attn"] = "on";
     QLOG_IF(
-        Priority::INFO,
-        "[LlamaModel] OpenCL backend selected: disabling flash attention by "
-        "default (not reliably supported on OpenCL)\n");
+        Priority::INFO, "[LlamaModel] Enabling flash attention by default\n");
   }
 
   constexpr int kAdrenoUbatchThreshold = 800;
@@ -213,22 +212,18 @@ void LlamaModel::tuneConfigMap(
     }
   }
 
-  // TurboQuant / PolarQuant KV-cache types (tbq3_0 / tbq4_0 / pq3_0 / pq4_0,
-  // ship Vulkan + CPU implementations only. On the OpenCL backend
-  // the kernels don't exist; on Metal the standalone MUL_MAT pipelines are
-  // explicitly disabled. Silently letting the user proceed lets llama.cpp
-  // commit KV-cache tensors to a backend that can't run the ops on them,
-  // which then aborts in ggml_backend_sched_split_graph at model load.
-  // Surface a clean error here instead.
-#if defined(__APPLE__)
-  constexpr bool kIsMetal = true;
-#else
-  constexpr bool kIsMetal = false;
-#endif
-  if (isOpenCl || kIsMetal) {
+  // Quantized KV-cache types are fragile on OpenCL: standard q-cache types can
+  // fail later during cache shifts, while TBQ/PQ kernels are not implemented.
+  // Surface a clean error here instead of letting llama.cpp commit KV-cache
+  // tensors to a backend that can't run the required ops.
+  if (isOpenCl || isMetal) {
     auto isTurboQuantKvType = [](const std::string& v) {
       return v == "tbq3_0" || v == "tbq4_0" || v == "pq3_0" || v == "pq4_0";
     };
+    auto isQuantizedKvType = [&](const std::string& v) {
+      return isTurboQuantKvType(v) || v == "q4_0" || v == "q4_1" ||
+             v == "q5_0" || v == "q5_1" || v == "q8_0" || v == "iq4_nl";
+    };
     auto checkCacheType = [&](const char* hyphenKey,
                               const char* underscoreKey,
                               const char* side) {
@@ -237,20 +232,30 @@ void LlamaModel::tuneConfigMap(
         it = configFilemap.find(underscoreKey);
       if (it == configFilemap.end())
         return;
-      if (!isTurboQuantKvType(it->second))
+      if (isOpenCl) {
+        if (!isQuantizedKvType(it->second))
+          return;
+      } else if (!isTurboQuantKvType(it->second)) {
         return;
+      }
       const char* backendName = isOpenCl ? "OpenCL" : "Metal";
+      const char* typeName = isTurboQuantKvType(it->second)
+                                 ? "TurboQuant/PolarQuant"
+                                 : "quantized";
+      const char* alternatives =
+          isOpenCl ? "f32/f16/bf16"
+                   : "f32/f16/bf16/q4_0/q4_1/q5_0/q5_1/q8_0/iq4_nl";
       throw qvac_errors::StatusError(
           qvac_errors::general_error::InvalidArgument,
           string_format(
-              "[LlamaModel] cache-type-%s=%s is a TurboQuant/PolarQuant type "
-              "and is not supported on the %s backend (Vulkan and CPU only). "
-              "Either pick a different cache type (f32/f16/bf16/q4_0/q4_1/"
-              "q5_0/q5_1/q8_0/iq4_nl) or switch device to a Vulkan GPU or "
-              "CPU.\n",
+              "[LlamaModel] cache-type-%s=%s is a %s KV-cache type and is not "
+              "supported on the %s backend. Either pick a different cache "
+              "type (%s) or switch device to a Vulkan GPU or CPU.\n",
               side,
               it->second.c_str(),
-              backendName));
+              typeName,
+              backendName,
+              alternatives));
     };
     checkCacheType("cache-type-k", "cache_type_k", "k");
     checkCacheType("cache-type-v", "cache_type_v", "v");
@@ -994,21 +999,23 @@ void LlamaModel::commonParamsParse(
         "embedded chat template is applied\n");
   }
 
-  // reasoning-budget controls whether the model emits a <think> reasoning
-  // channel. -1 (default) leaves it on; 0 disables. `std::from_chars` is used
-  // instead of `std::stoi` because the latter accepts trailing garbage ("0abc"
-  // → 0) and throws an uncaught `std::out_of_range` on overflow.
+  // reasoning-budget controls the size of the model's <think> reasoning
+  // channel: -1 (default) leaves it unrestricted, 0 disables thinking
+  // entirely, any positive N caps the reasoning channel at N tokens (the
+  // budget sampler forces </think> once N reasoning tokens have been
+  // emitted).
   auto parseReasoningBudget = [](const std::string& raw) {
     int value = 0;
     const char* begin = raw.data();
     const char* end = begin + raw.size();
     const auto [ptr, ec] = std::from_chars(begin, end, value);
-    if (ec != std::errc{} || ptr != end || (value != 0 && value != -1)) {
+    if (ec != std::errc{} || ptr != end || value < -1) {
       throw qvac_errors::StatusError(
           ADDON_ID,
           qvac_errors::general_error::toString(
               qvac_errors::general_error::InvalidArgument),
-          "reasoning-budget must be -1 (unrestricted) or 0 (disabled)");
+          "reasoning-budget must be -1 (unrestricted), 0 (disabled), or a "
+          "positive integer (token cap)");
     }
     return value;
   };
@@ -1114,6 +1121,7 @@ void LlamaModel::commonParamsParse(
   }
 
   bool isOpenCl = false;
+  bool isMetal = false;
   {
     using namespace backend_selection;
     const BackendType preferredBackend =
@@ -1180,6 +1188,9 @@ void LlamaModel::commonParamsParse(
 
     isOpenCl = chosenBackend.first == BackendType::GPU &&
                chosenBackend.second.find("opencl") != std::string::npos;
+    isMetal = chosenBackend.first == BackendType::GPU &&
+              (chosenBackend.second.find("metal") != std::string::npos ||
+               chosenBackend.second.rfind("mtl", 0) == 0);
   }
 
   tuneConfigMap(
@@ -1187,7 +1198,8 @@ void LlamaModel::commonParamsParse(
       metadata_,
       outAdrenoVersion,
       pendingFinetuneOverrides_,
-      isOpenCl);
+      isOpenCl,
+      isMetal);
 
   // Handle both reverse-prompt variants
   for (const std::string& key : {"reverse-prompt", "reverse_prompt"}) {
@@ -1310,10 +1322,6 @@ void LlamaModel::commonParamsParse(
   postprocess_cpu_params(params.cpuparams, nullptr);
   postprocess_cpu_params(params.cpuparams_batch, &params.cpuparams);
 
-  postprocess_cpu_params(params.speculative.draft.cpuparams, &params.cpuparams);
-  postprocess_cpu_params(
-      params.speculative.draft.cpuparams_batch, &params.cpuparams_batch);
-
   if (!params.kv_overrides.empty()) {
     params.kv_overrides.emplace_back();
     params.kv_overrides.back().key[0] = 0;
diff --git a/packages/llm-llamacpp/addon/src/model-interface/LlamaModel.hpp b/packages/llm-llamacpp/addon/src/model-interface/LlamaModel.hpp
index adb93e8a1b..a7ef25e507 100644
--- a/packages/llm-llamacpp/addon/src/model-interface/LlamaModel.hpp
+++ b/packages/llm-llamacpp/addon/src/model-interface/LlamaModel.hpp
@@ -67,14 +67,15 @@ class LlamaModel : public IModel, public IModelAsyncLoad, public IModelCancel {
   /// @param adrenoVersion Detected Adreno GPU version, if any.
   /// @param finetuneOverrides If set, finetuning mode is active with these
   /// context/batch params and GPU caps.
-  /// @param isOpenCl True when the chosen GPU backend is OpenCL; used to
-  /// disable flash-attn by default since it is not reliably supported on
-  /// the OpenCL backend.
+  /// @param isOpenCl True when the chosen GPU backend is OpenCL; used to reject
+  /// unsupported quantized KV-cache types.
+  /// @param isMetal True when the chosen GPU backend is Metal; used to reject
+  /// unsupported TurboQuant/PolarQuant KV-cache types.
   static void tuneConfigMap(
       std::unordered_map<std::string, std::string>& configFilemap,
       const ModelMetaData& metadata, const std::optional<int>& adrenoVersion,
       const FinetuneConfigOverrides& finetuneOverrides = {},
-      bool isOpenCl = false);
+      bool isOpenCl = false, bool isMetal = false);
 
   /**
    * The Constructor for llama model.
diff --git a/packages/llm-llamacpp/addon/src/model-interface/LlmContext.hpp b/packages/llm-llamacpp/addon/src/model-interface/LlmContext.hpp
index 2b18b1e030..474c24b9b7 100644
--- a/packages/llm-llamacpp/addon/src/model-interface/LlmContext.hpp
+++ b/packages/llm-llamacpp/addon/src/model-interface/LlmContext.hpp
@@ -34,10 +34,11 @@ struct GenerationParams {
   // Mutually exclusive with `grammar` — the JS wrapper rejects requests
   // that set both. Mirrors the load-time `--json-schema` flag.
   std::optional<std::string> json_schema;
-  // Reasoning channel budget override. `-1` keeps reasoning on, `0` disables
-  // it for this request. Mirrors the load-time `reasoning-budget` config; the
-  // override is applied to `params_.reasoning_budget` for the duration of the
-  // request and restored on completion.
+  // Reasoning channel budget override. `-1` keeps reasoning unrestricted, `0`
+  // disables it, and positive values cap the reasoning channel at that many
+  // tokens. Mirrors the load-time `reasoning-budget` config; the override is
+  // applied to `params_.reasoning_budget` for the duration of the request and
+  // restored on completion.
   std::optional<int> reasoning_budget;
 
   [[nodiscard]] bool hasOverrides() const {
diff --git a/packages/llm-llamacpp/addon/src/model-interface/MtmdLlmContext.cpp b/packages/llm-llamacpp/addon/src/model-interface/MtmdLlmContext.cpp
index 788d376c71..9e59e5cdc6 100644
--- a/packages/llm-llamacpp/addon/src/model-interface/MtmdLlmContext.cpp
+++ b/packages/llm-llamacpp/addon/src/model-interface/MtmdLlmContext.cpp
@@ -216,7 +216,20 @@ void MtmdLlmContext::tokenizeChat(
   if (!tools.empty()) {
     inputs.tools = tools;
   }
-  formattedChat = getPrompt(tmpls_.get(), inputs, &thinkingForcedOpen_);
+  std::string thinkingStartTag;
+  std::string thinkingEndTag;
+  std::string generationPrompt;
+  formattedChat = getPrompt(
+      tmpls_.get(),
+      inputs,
+      &thinkingForcedOpen_,
+      &thinkingStartTag,
+      &thinkingEndTag,
+      &generationPrompt);
+  thinkingForcedOpenText_ =
+      thinkingForcedOpen_
+          ? getThinkingForcedOpenText(generationPrompt, thinkingStartTag)
+          : std::string{};
 
   if (formattedChat.empty()) {
     std::string errorMsg = string_format(
@@ -224,6 +237,21 @@ void MtmdLlmContext::tokenizeChat(
     throw qvac_errors::StatusError(ADDON_ID, toString(EmptyPrompt), errorMsg);
   }
 
+  if (configureReasoningBudgetSampling(
+          params_,
+          modelCtx_.lctx,
+          thinkingStartTag,
+          thinkingEndTag,
+          generationPrompt)) {
+    smpl_.reset(common_sampler_init(modelCtx_.model, params_.sampling));
+    if (!smpl_) {
+      std::string errorMsg = string_format(
+          "[MtmdLlm] %s: failed to initialize sampling subsystem\n", __func__);
+      throw qvac_errors::StatusError(
+          ADDON_ID, toString(UnableToCreateSamplingSystem), errorMsg);
+    }
+  }
+
   QLOG_IF(
       Priority::DEBUG,
       string_format("[MtmdLlm] formatted prompt: %s\n", formattedChat.c_str()));
@@ -492,10 +520,10 @@ bool MtmdLlmContext::generateResponse(
   if (thinkingForcedOpen_ && outputCallback) {
     // MtmdLlmContext doesn't carry a reasoningState_ (no reasoning-aware EOS
     // replacement on the multimodal path today), so unlike TextLlmContext we
-    // only prepend the visible "<think>\n" opener and don't flip an
+    // only prepend the visible reasoning opener and don't flip an
     // inside_reasoning flag. If reasoning state is added here later, mirror
     // TextLlmContext::generateResponse and set it true alongside this emit.
-    outputCallback("<think>\n");
+    outputCallback(thinkingForcedOpenText_);
   }
 
   if (stopGeneration_.load()) {
@@ -717,6 +745,8 @@ void MtmdLlmContext::resetState(bool resetStats) {
 
   // Clear UTF-8 buffer when resetting state
   utf8Buffer_.clear();
+  thinkingForcedOpen_ = false;
+  thinkingForcedOpenText_.clear();
 
   clearSequenceMemory(modelCtx_.lctx);
 
diff --git a/packages/llm-llamacpp/addon/src/model-interface/MtmdLlmContext.hpp b/packages/llm-llamacpp/addon/src/model-interface/MtmdLlmContext.hpp
index 356743c796..dcca7c7379 100644
--- a/packages/llm-llamacpp/addon/src/model-interface/MtmdLlmContext.hpp
+++ b/packages/llm-llamacpp/addon/src/model-interface/MtmdLlmContext.hpp
@@ -230,10 +230,11 @@ class MtmdLlmContext : public LlmContext {
   bool isHarmonyModel_ = false;
   llama_token harmonyCallToken_ = LLAMA_TOKEN_NULL;
 
-  // Force-opens the reasoning channel in the prompt suffix to prepend the
-  // matching "<think>\n" opener to the visible stream so consumers see balanced
+  // Force-opens the reasoning channel in the prompt suffix. The text mirrors
+  // the template-specific visible reasoning opener so consumers see balanced
   // tags.
   bool thinkingForcedOpen_ = false;
+  std::string thinkingForcedOpenText_;
 
   std::atomic<bool> stopGeneration_ = false;
 };
diff --git a/packages/llm-llamacpp/addon/src/model-interface/TextLlmContext.cpp b/packages/llm-llamacpp/addon/src/model-interface/TextLlmContext.cpp
index e7aa4860ee..7c9a60e230 100644
--- a/packages/llm-llamacpp/addon/src/model-interface/TextLlmContext.cpp
+++ b/packages/llm-llamacpp/addon/src/model-interface/TextLlmContext.cpp
@@ -284,7 +284,34 @@ void TextLlmContext::tokenizeChat(
   if (!tools.empty()) {
     inputs.tools = tools;
   }
-  prompt = getPrompt(tmpls_.get(), inputs, &thinkingForcedOpen_);
+  std::string thinkingStartTag;
+  std::string thinkingEndTag;
+  std::string generationPrompt;
+  prompt = getPrompt(
+      tmpls_.get(),
+      inputs,
+      &thinkingForcedOpen_,
+      &thinkingStartTag,
+      &thinkingEndTag,
+      &generationPrompt);
+  thinkingForcedOpenText_ =
+      thinkingForcedOpen_
+          ? getThinkingForcedOpenText(generationPrompt, thinkingStartTag)
+          : std::string{};
+  if (configureReasoningBudgetSampling(
+          params_,
+          modelCtx_.lctx,
+          thinkingStartTag,
+          thinkingEndTag,
+          generationPrompt)) {
+    smpl_.reset(common_sampler_init(modelCtx_.model, params_.sampling));
+    if (!smpl_) {
+      std::string errorMsg = string_format(
+          "[TextLlm] %s: failed to initialize sampling subsystem\n", __func__);
+      throw qvac_errors::StatusError(
+          ADDON_ID, toString(UnableToCreateSamplingSystem), errorMsg);
+    }
+  }
 
   QLOG_IF(
       Priority::DEBUG,
@@ -598,12 +625,10 @@ bool TextLlmContext::generateResponse(
   assistantOutput_.clear();
   generationStarted_ = false;
 
-  // The chat template force-opened the reasoning channel in the prompt (e.g.
-  // Qwen3-style / DeepSeek-R1 templates end with "<think>\n"), so the model
-  // resumes generation INSIDE the reasoning block. (Gemma4's reasoning channel
-  // is model-emitted upstream and does not set this flag.)
+  // The chat template force-opened the reasoning channel in the prompt, so the
+  // model resumes generation inside the reasoning block.
   if (thinkingForcedOpen_ && outputCallback) {
-    outputCallback("<think>\n");
+    outputCallback(thinkingForcedOpenText_);
     reasoningState_.inside_reasoning = true;
   }
 
@@ -899,6 +924,8 @@ void TextLlmContext::resetState(bool resetStats) {
   forcedTokens_.clear();
   assistantOutput_.clear();
   generationStarted_ = false;
+  thinkingForcedOpen_ = false;
+  thinkingForcedOpenText_.clear();
 
   clearSequenceMemory(modelCtx_.lctx);
 
diff --git a/packages/llm-llamacpp/addon/src/model-interface/TextLlmContext.hpp b/packages/llm-llamacpp/addon/src/model-interface/TextLlmContext.hpp
index d53764c9c0..46d3068a64 100644
--- a/packages/llm-llamacpp/addon/src/model-interface/TextLlmContext.hpp
+++ b/packages/llm-llamacpp/addon/src/model-interface/TextLlmContext.hpp
@@ -270,10 +270,11 @@ class TextLlmContext : public LlmContext, public SequenceDriver {
   bool isHarmonyModel_ = false;
   llama_token harmonyCallToken_ = LLAMA_TOKEN_NULL;
 
-  // Force-opens the reasoning channel in the prompt suffix to prepend the
-  // matching "<think>\n" opener to the visible stream so consumers see balanced
+  // Force-opens the reasoning channel in the prompt suffix. The text mirrors
+  // the template-specific visible reasoning opener so consumers see balanced
   // tags.
   bool thinkingForcedOpen_ = false;
+  std::string thinkingForcedOpenText_;
 
   std::atomic<bool> stopGeneration_ = false;
 };
diff --git a/packages/llm-llamacpp/addon/src/utils/ChatTemplateUtils.cpp b/packages/llm-llamacpp/addon/src/utils/ChatTemplateUtils.cpp
index 9bc66778b5..3e06c1c249 100644
--- a/packages/llm-llamacpp/addon/src/utils/ChatTemplateUtils.cpp
+++ b/packages/llm-llamacpp/addon/src/utils/ChatTemplateUtils.cpp
@@ -4,6 +4,7 @@
 #include <cctype>
 #include <ranges>
 #include <string_view>
+#include <utility>
 
 #include <llama.h>
 
@@ -174,11 +175,22 @@ std::string getChatTemplate(
 
 std::string getPrompt(
     const struct common_chat_templates* tmpls,
-    struct common_chat_templates_inputs& inputs, bool* outThinkingForcedOpen) {
+    struct common_chat_templates_inputs& inputs, bool* outThinkingForcedOpen,
+    std::string* outThinkingStartTag, std::string* outThinkingEndTag,
+    std::string* outGenerationPrompt) {
   auto exportParams = [&](const common_chat_params& params) {
     if (outThinkingForcedOpen) {
       *outThinkingForcedOpen = params.thinking_forced_open;
     }
+    if (outThinkingStartTag) {
+      *outThinkingStartTag = params.thinking_start_tag;
+    }
+    if (outThinkingEndTag) {
+      *outThinkingEndTag = params.thinking_end_tag;
+    }
+    if (outGenerationPrompt) {
+      *outGenerationPrompt = params.generation_prompt;
+    }
   };
   try {
     auto params = common_chat_templates_apply(tmpls, inputs);
@@ -211,5 +223,57 @@ std::string getPrompt(
   }
 }
 
+bool configureReasoningBudgetSampling(
+    common_params& params, ::llama_context* lctx,
+    const std::string& thinkingStartTag, const std::string& thinkingEndTag,
+    const std::string& generationPrompt) {
+  common_params_sampling next = params.sampling;
+  next.reasoning_budget_tokens =
+      params.reasoning_budget > 0 ? params.reasoning_budget : -1;
+  next.reasoning_budget_start.clear();
+  next.reasoning_budget_end.clear();
+  next.reasoning_budget_forced.clear();
+  next.generation_prompt.clear();
+
+  if (params.reasoning_budget > 0 && lctx != nullptr &&
+      !thinkingEndTag.empty()) {
+    next.generation_prompt = generationPrompt;
+    if (!thinkingStartTag.empty()) {
+      next.reasoning_budget_start =
+          common_tokenize(lctx, thinkingStartTag, false, true);
+    }
+    next.reasoning_budget_end =
+        common_tokenize(lctx, thinkingEndTag, false, true);
+    next.reasoning_budget_forced = common_tokenize(
+        lctx,
+        params.sampling.reasoning_budget_message + thinkingEndTag,
+        false,
+        true);
+  }
+
+  const bool changed =
+      params.sampling.reasoning_budget_tokens != next.reasoning_budget_tokens ||
+      params.sampling.reasoning_budget_start != next.reasoning_budget_start ||
+      params.sampling.reasoning_budget_end != next.reasoning_budget_end ||
+      params.sampling.reasoning_budget_forced != next.reasoning_budget_forced ||
+      params.sampling.generation_prompt != next.generation_prompt;
+  if (changed) {
+    params.sampling = std::move(next);
+  }
+  return changed;
+}
+
+std::string getThinkingForcedOpenText(
+    const std::string& generationPrompt, const std::string& thinkingStartTag) {
+  if (thinkingStartTag.empty()) {
+    return {};
+  }
+  const auto start = generationPrompt.rfind(thinkingStartTag);
+  if (start == std::string::npos) {
+    return thinkingStartTag;
+  }
+  return generationPrompt.substr(start);
+}
+
 } // namespace utils
 } // namespace qvac_lib_inference_addon_llama
diff --git a/packages/llm-llamacpp/addon/src/utils/ChatTemplateUtils.hpp b/packages/llm-llamacpp/addon/src/utils/ChatTemplateUtils.hpp
index b523e51b98..3652d91464 100644
--- a/packages/llm-llamacpp/addon/src/utils/ChatTemplateUtils.hpp
+++ b/packages/llm-llamacpp/addon/src/utils/ChatTemplateUtils.hpp
@@ -71,11 +71,35 @@ std::string getChatTemplate(
  *
  * @p outThinkingForcedOpen (optional) receives the flag indicating that the
  *    template force-opened the reasoning channel in the prompt suffix.
+ * @p outThinkingStartTag (optional) receives the template-specific reasoning
+ *    start tag, when the template exposes one.
+ * @p outThinkingEndTag (optional) receives the template-specific reasoning
+ *    end tag, when the template exposes one.
+ * @p outGenerationPrompt (optional) receives the assistant generation prompt
+ *    already appended to the formatted prompt.
  */
 std::string getPrompt(
     const struct common_chat_templates* tmpls,
     struct common_chat_templates_inputs& inputs,
-    bool* outThinkingForcedOpen = nullptr);
+    bool* outThinkingForcedOpen = nullptr,
+    std::string* outThinkingStartTag = nullptr,
+    std::string* outThinkingEndTag = nullptr,
+    std::string* outGenerationPrompt = nullptr);
+
+/**
+ * @brief Configures the common-sampling reasoning-budget fields from
+ * template-derived thinking tags.
+ *
+ * Returns true when the sampling block changed and the caller should recreate
+ * the common_sampler.
+ */
+bool configureReasoningBudgetSampling(
+    common_params& params, ::llama_context* lctx,
+    const std::string& thinkingStartTag, const std::string& thinkingEndTag,
+    const std::string& generationPrompt);
+
+std::string getThinkingForcedOpenText(
+    const std::string& generationPrompt, const std::string& thinkingStartTag);
 
 } // namespace utils
 } // namespace qvac_lib_inference_addon_llama
diff --git a/packages/llm-llamacpp/index.d.ts b/packages/llm-llamacpp/index.d.ts
index 95abd62d73..6d0caa71cd 100644
--- a/packages/llm-llamacpp/index.d.ts
+++ b/packages/llm-llamacpp/index.d.ts
@@ -74,8 +74,13 @@ export interface LlamaConfig {
   'cache-type-v'?: string
   /** Writable directory for OpenCL kernel binary cache. Required on Android for fast GPU startup. */
   openclCacheDir?: string
-  /** Reasoning channel budget. `-1` (default) leaves the model's reasoning channel on; `0` disables it. */
-  reasoning_budget?: -1 | 0 | '-1' | '0'
+  /**
+   * Reasoning channel budget. `-1` (default) leaves the model's reasoning
+   * channel on; `0` disables it; any positive integer caps the reasoning
+   * channel at that many tokens (the sampler force-emits `</think>` once
+   * the budget is exhausted).
+   */
+  reasoning_budget?: number | `${number}`
   /**
    * Number of concurrent sequence slots for continuous-batching (`--parallel` /
    * `n_parallel` in llama.cpp). Values `>= 2` activate the continuous-batch
@@ -157,11 +162,12 @@ export interface GenerationParams {
   json_schema?: string | Record<string, unknown>
   /**
    * Per-request reasoning channel budget. `-1` keeps the model's reasoning
-   * channel on; `0` disables it for this request. Equivalent to the load-time
+   * channel on; `0` disables it for this request; any positive integer caps
+   * the reasoning channel at that many tokens. Equivalent to the load-time
    * `reasoning_budget` config but scoped to a single `run()` call; the prior
    * value is restored afterwards.
    */
-  reasoning_budget?: -1 | 0
+  reasoning_budget?: number
 }
 
 export interface RunOptions {
diff --git a/packages/llm-llamacpp/scripts/generate-mobile-integration-tests.js b/packages/llm-llamacpp/scripts/generate-mobile-integration-tests.js
index 87f1cbec2f..3560e2c6bb 100644
--- a/packages/llm-llamacpp/scripts/generate-mobile-integration-tests.js
+++ b/packages/llm-llamacpp/scripts/generate-mobile-integration-tests.js
@@ -9,9 +9,6 @@ const integrationDir = path.join(repoRoot, 'test', 'integration')
 const mobileDir = path.join(repoRoot, 'test', 'mobile')
 const outputFile = path.join(mobileDir, 'integration.auto.cjs')
 const groupsFile = path.join(mobileDir, 'test-groups.json')
-const mobileExcludedTests = new Set([
-  'continuous-batching.test.js'
-])
 
 // The benchmark-perf-*.test.js shards are generated, not committed (see
 // .gitignore), but the committed integration.auto.cjs references them. Enumerating
@@ -40,7 +37,6 @@ function getIntegrationFiles () {
 
   return fs.readdirSync(integrationDir)
     .filter(entry => entry.endsWith('.test.js'))
-    .filter(entry => !mobileExcludedTests.has(entry))
     .sort()
 }
 
diff --git a/packages/llm-llamacpp/test/integration/_vlm-image-perf.js b/packages/llm-llamacpp/test/integration/_vlm-image-perf.js
index 268a2e51ac..a11295c1fe 100644
--- a/packages/llm-llamacpp/test/integration/_vlm-image-perf.js
+++ b/packages/llm-llamacpp/test/integration/_vlm-image-perf.js
@@ -36,12 +36,11 @@ function _envInt (key, fallback) {
 const PERF_RUNS = _envInt('QVAC_PERF_RUNS', 1)
 const PERF_WARMUP_RUNS = _envInt('QVAC_PERF_WARMUP_RUNS', 1)
 
-// QVAC-19368: skip the heaviest image (aurora) only on Android non-benchmark
-// on-PR runs where the 30-min Device Farm per-test cap is tight. iOS and
-// desktop always run aurora (fast enough). The benchmark (QVAC_PERF_ONLY=true)
-// always runs the full set regardless of platform. Uses the explicit
-// QVAC_PERF_ONLY flag (already plumbed to the device via the testspec config)
-// instead of proxying off PERF_RUNS, per review feedback.
+// QVAC-19368: skip the heaviest image (aurora) on Android non-benchmark
+// on-PR runs where the 30-min Device Farm per-test cap is tight. The
+// benchmark (QVAC_PERF_ONLY=true) runs the full set on supported platforms.
+// Uses the explicit QVAC_PERF_ONLY flag (already plumbed to the device via
+// the testspec config) instead of proxying off PERF_RUNS, per review feedback.
 function _envStr (key) {
   if (typeof os.getEnv === 'function') return os.getEnv(key) || ''
   if (typeof process !== 'undefined' && process.env) return process.env[key] || ''
@@ -216,6 +215,7 @@ module.exports = {
   IMAGE_CASES,
   GEMMA4_MODEL,
   QWEN35_MODEL,
+  isDarwinX64,
   skipHeavyImages,
   runVlmImagePerf
 }
diff --git a/packages/llm-llamacpp/test/integration/cache-state-machine.test.js b/packages/llm-llamacpp/test/integration/cache-state-machine.test.js
index 9e919270e5..6cdc37b706 100644
--- a/packages/llm-llamacpp/test/integration/cache-state-machine.test.js
+++ b/packages/llm-llamacpp/test/integration/cache-state-machine.test.js
@@ -3,7 +3,7 @@
 const path = require('bare-path')
 const fs = require('bare-fs')
 const LlmLlamacpp = require('../../index.js')
-const { cleanupIntegrationCacheFiles, ensureModel, safeTest } = require('./utils')
+const { cleanupIntegrationCacheFiles, ensureModel, safeTest: integrationTest } = require('./utils')
 const { attachSpecLogger } = require('./spec-logger')
 const os = require('bare-os')
 
@@ -11,6 +11,10 @@ const isDarwinX64 = os.platform() === 'darwin' && os.arch() === 'x64'
 const isLinuxArm64 = os.platform() === 'linux' && os.arch() === 'arm64'
 const useCpu = isDarwinX64 || isLinuxArm64
 
+function safeTest (name, opts, fn) {
+  integrationTest(name, { ...opts, skip: opts.skip || isDarwinX64 }, fn)
+}
+
 const DEFAULT_MODEL = {
   name: 'Llama-3.2-1B-Instruct-Q4_0.gguf',
   url: 'https://huggingface.co/bartowski/Llama-3.2-1B-Instruct-GGUF/resolve/main/Llama-3.2-1B-Instruct-Q4_0.gguf'
diff --git a/packages/llm-llamacpp/test/integration/continuous-batching.test.js b/packages/llm-llamacpp/test/integration/continuous-batching.test.js
index 2612c3b333..9928e7081e 100644
--- a/packages/llm-llamacpp/test/integration/continuous-batching.test.js
+++ b/packages/llm-llamacpp/test/integration/continuous-batching.test.js
@@ -1,11 +1,10 @@
 'use strict'
 
-const test = require('brittle')
 const path = require('bare-path')
 const os = require('bare-os')
 const process = require('bare-process')
 const LlmLlamacpp = require('../../index.js')
-const { ensureModel } = require('./utils')
+const { ensureModel, safeTest } = require('./utils')
 const { attachSpecLogger } = require('./spec-logger')
 
 const platform = os.platform()
@@ -152,7 +151,7 @@ function logStreamingProgress (response) {
 // 1B throughput/correctness run is too slow or complicated for mobile and legacy macOS x64.
 const skipHeavyPlatform = isMobile || isDarwin
 
-test('continuous batching answers 16 prompts correctly and improves Linux GPU TPS', { timeout: 900_000, skip: skipHeavyPlatform }, async t => {
+safeTest('continuous batching answers 16 prompts correctly and improves Linux GPU TPS', { timeout: 900_000, skip: skipHeavyPlatform }, async t => {
   const singleModel = await setupModel(t)
   const singleNativeTpsValues = []
   const singleWallTpsValues = []
diff --git a/packages/llm-llamacpp/test/integration/gemma4-image-elephant-perf.test.js b/packages/llm-llamacpp/test/integration/gemma4-image-elephant-perf.test.js
index eac37c3246..8bd2113ba3 100644
--- a/packages/llm-llamacpp/test/integration/gemma4-image-elephant-perf.test.js
+++ b/packages/llm-llamacpp/test/integration/gemma4-image-elephant-perf.test.js
@@ -4,9 +4,9 @@
 // 30-minute mobile cap. Asserts the elephant keyword + records perf.
 
 const test = require('brittle')
-const { GEMMA4_MODEL, IMAGE_CASES, runVlmImagePerf } = require('./_vlm-image-perf.js')
+const { GEMMA4_MODEL, IMAGE_CASES, isDarwinX64, runVlmImagePerf } = require('./_vlm-image-perf.js')
 
-test('Gemma4-VL image perf [elephant]', { timeout: 1_800_000 }, async t => {
+test('Gemma4-VL image perf [elephant]', { timeout: 1_800_000, skip: isDarwinX64 }, async t => {
   await runVlmImagePerf(t, GEMMA4_MODEL, IMAGE_CASES.elephant)
 })
 
diff --git a/packages/llm-llamacpp/test/integration/gemma4-image-fruit-plate-perf.test.js b/packages/llm-llamacpp/test/integration/gemma4-image-fruit-plate-perf.test.js
index 1e6f048f0b..8c695d460a 100644
--- a/packages/llm-llamacpp/test/integration/gemma4-image-fruit-plate-perf.test.js
+++ b/packages/llm-llamacpp/test/integration/gemma4-image-fruit-plate-perf.test.js
@@ -4,9 +4,9 @@
 // 30-minute mobile cap. Asserts a fruit keyword + records perf.
 
 const test = require('brittle')
-const { GEMMA4_MODEL, IMAGE_CASES, runVlmImagePerf } = require('./_vlm-image-perf.js')
+const { GEMMA4_MODEL, IMAGE_CASES, isDarwinX64, runVlmImagePerf } = require('./_vlm-image-perf.js')
 
-test('Gemma4-VL image perf [fruit plate]', { timeout: 1_800_000 }, async t => {
+test('Gemma4-VL image perf [fruit plate]', { timeout: 1_800_000, skip: isDarwinX64 }, async t => {
   await runVlmImagePerf(t, GEMMA4_MODEL, IMAGE_CASES['fruit-plate'])
 })
 
diff --git a/packages/llm-llamacpp/test/integration/gemma4-image-high-res-aurora-perf.test.js b/packages/llm-llamacpp/test/integration/gemma4-image-high-res-aurora-perf.test.js
index 0d05133b8b..ae0666843d 100644
--- a/packages/llm-llamacpp/test/integration/gemma4-image-high-res-aurora-perf.test.js
+++ b/packages/llm-llamacpp/test/integration/gemma4-image-high-res-aurora-perf.test.js
@@ -4,12 +4,12 @@
 // the 30-minute mobile cap. Asserts an aurora keyword + records perf.
 
 const test = require('brittle')
-const { GEMMA4_MODEL, IMAGE_CASES, skipHeavyImages, runVlmImagePerf } = require('./_vlm-image-perf.js')
+const { GEMMA4_MODEL, IMAGE_CASES, isDarwinX64, skipHeavyImages, runVlmImagePerf } = require('./_vlm-image-perf.js')
 
 // QVAC-19368: aurora is the heaviest image; skip it on Android on-PR runs
-// where the 30-min Device Farm cap is tight. iOS + desktop always run it.
-// The benchmark (QVAC_PERF_ONLY=true) runs all 3 images on all platforms.
-test('Gemma4-VL image perf [high-res aurora]', { timeout: 1_800_000, skip: skipHeavyImages }, async t => {
+// where the 30-min Device Farm cap is tight. Darwin x64 is skipped separately.
+// The benchmark (QVAC_PERF_ONLY=true) runs all 3 images on supported platforms.
+test('Gemma4-VL image perf [high-res aurora]', { timeout: 1_800_000, skip: isDarwinX64 || skipHeavyImages }, async t => {
   await runVlmImagePerf(t, GEMMA4_MODEL, IMAGE_CASES['high-res-aurora'])
 })
 
diff --git a/packages/llm-llamacpp/test/integration/model-loading.test.js b/packages/llm-llamacpp/test/integration/model-loading.test.js
index c1f084450d..989500b3a5 100644
--- a/packages/llm-llamacpp/test/integration/model-loading.test.js
+++ b/packages/llm-llamacpp/test/integration/model-loading.test.js
@@ -13,6 +13,7 @@ const arch = os.arch()
 const isDarwinX64 = platform === 'darwin' && arch === 'x64'
 const isLinuxArm64 = platform === 'linux' && arch === 'arm64'
 const isLinuxX64 = platform === 'linux' && arch === 'x64'
+const isWindowsX64 = platform === 'win32' && arch === 'x64'
 const useCpu = isDarwinX64 || isLinuxArm64
 
 const DEFAULT_MODEL = {
@@ -126,10 +127,9 @@ const SHARDED_MODEL = {
   baseUrl: 'https://huggingface.co/jmb95/Qwen3-0.6B-UD-IQ1_S-sharded/resolve/main/'
 }
 
-// This test can take longer to download and execute. To avoid blowing up testing time on all
-// platforms, just use Linux for now. C++ tests already have faster coverage for each type
-// of load.
-test('sharded model can run inference end-to-end', { timeout: 4 * 60 * 1000, skip: !isLinuxX64 }, async t => {
+// This test can take longer to download and execute. Keep it on desktop x64
+// runners where the sharded loader has CI coverage.
+test('sharded model can run inference end-to-end', { timeout: 4 * 60 * 1000, skip: !(isLinuxX64 || isWindowsX64) }, async t => {
   const fs = require('bare-fs')
   const modelDir = path.resolve(__dirname, '../model')
   fs.mkdirSync(modelDir, { recursive: true })
diff --git a/packages/llm-llamacpp/test/integration/mrope-sliding-context.test.js b/packages/llm-llamacpp/test/integration/mrope-sliding-context.test.js
index 7862c0b5bd..ae9a8730ea 100644
--- a/packages/llm-llamacpp/test/integration/mrope-sliding-context.test.js
+++ b/packages/llm-llamacpp/test/integration/mrope-sliding-context.test.js
@@ -4,7 +4,7 @@ const fs = require('bare-fs')
 const os = require('bare-os')
 const path = require('bare-path')
 const LlmLlamacpp = require('../../index.js')
-const { cleanupIntegrationCacheFiles, ensureModel, getMediaPath, safeTest } = require('./utils')
+const { cleanupIntegrationCacheFiles, ensureModel, getMediaPath, safeTest: integrationTest } = require('./utils')
 
 const platform = os.platform()
 const arch = os.arch()
@@ -16,6 +16,10 @@ const isLinuxArm64 = platform === 'linux' && arch === 'arm64'
 const useCpu = isDarwinX64 || isLinuxArm64
 const skipTbqPq = isDarwin || isIos || isAndroid
 
+function safeTest (name, opts, fn) {
+  integrationTest(name, { ...opts, skip: opts.skip || isDarwinX64 }, fn)
+}
+
 const QWEN3_5_MODEL = {
   name: 'Qwen3.5-0.8B-Q8_0.gguf',
   url: 'https://huggingface.co/unsloth/Qwen3.5-0.8B-GGUF/resolve/main/Qwen3.5-0.8B-Q8_0.gguf'
diff --git a/packages/llm-llamacpp/test/integration/multi-gpu.test.js b/packages/llm-llamacpp/test/integration/multi-gpu.test.js
index d6a926370e..21ea45b819 100644
--- a/packages/llm-llamacpp/test/integration/multi-gpu.test.js
+++ b/packages/llm-llamacpp/test/integration/multi-gpu.test.js
@@ -1,11 +1,14 @@
 'use strict'
 
 const process = require('bare-process')
+const os = require('bare-os')
 const LlmLlamacpp = require('../../index.js')
 const { ensureModel, safeTest } = require('./utils')
 const { attachSpecLogger } = require('./spec-logger')
 const path = require('bare-path')
 
+const isDarwinX64 = os.platform() === 'darwin' && os.arch() === 'x64'
+
 const MODEL = {
   name: 'Qwen3-0.6B-Q8_0.gguf',
   url: 'https://huggingface.co/unsloth/Qwen3-0.6B-GGUF/resolve/main/Qwen3-0.6B-Q8_0.gguf'
@@ -32,6 +35,7 @@ async function collectResponse (response) {
 }
 
 const hasMultiGpu = process.env.QVAC_HAS_MULTI_GPU === '1'
+const skip = isDarwinX64
 
 const BASE_CONFIG = {
   device: 'gpu',
@@ -92,19 +96,19 @@ function assertSingleDevice (t, devices) {
   t.ok(devices.size <= 1, `layers should stay on a single device (found: ${[...devices].join(', ')})`)
 }
 
-safeTest('multi-gpu: split-mode=layer distributes layers across GPUs', { timeout: 600_000 }, async t => {
+safeTest('multi-gpu: split-mode=layer distributes layers across GPUs', { timeout: 600_000, skip }, async t => {
   await runMultiGpuTest(t, { 'split-mode': 'layer' }, assertMultiDevice('layers'))
 })
 
-safeTest('multi-gpu: split-mode=row distributes tensors across GPUs', { timeout: 600_000 }, async t => {
+safeTest('multi-gpu: split-mode=row distributes tensors across GPUs', { timeout: 600_000, skip }, async t => {
   await runMultiGpuTest(t, { 'split-mode': 'row' }, assertMultiDevice('tensors'))
 })
 
-safeTest('multi-gpu: default (no split-mode) pins layers to a single device', { timeout: 600_000 }, async t => {
+safeTest('multi-gpu: default (no split-mode) pins layers to a single device', { timeout: 600_000, skip }, async t => {
   await runMultiGpuTest(t, {}, assertSingleDevice)
 })
 
-safeTest('multi-gpu: split-mode=layer with tensor-split and main-gpu', { timeout: 600_000 }, async t => {
+safeTest('multi-gpu: split-mode=layer with tensor-split and main-gpu', { timeout: 600_000, skip }, async t => {
   await runMultiGpuTest(
     t,
     { 'split-mode': 'layer', 'tensor-split': '1,1', 'main-gpu': '0' },
diff --git a/packages/llm-llamacpp/test/integration/multi-instance.test.js b/packages/llm-llamacpp/test/integration/multi-instance.test.js
index cc458d9486..1d958cbdbf 100644
--- a/packages/llm-llamacpp/test/integration/multi-instance.test.js
+++ b/packages/llm-llamacpp/test/integration/multi-instance.test.js
@@ -9,8 +9,8 @@ const platform = os.platform()
 const arch = os.arch()
 const isDarwinX64 = platform === 'darwin' && arch === 'x64'
 const isLinuxArm64 = platform === 'linux' && arch === 'arm64'
-const isWindowsX64 = platform === 'win32' && arch === 'x64'
 const useCpu = isDarwinX64 || isLinuxArm64
+const skip = isDarwinX64
 
 const DEFAULT_MODEL = {
   name: 'Llama-3.2-1B-Instruct-Q4_0.gguf',
@@ -73,7 +73,7 @@ async function collectResponse (response) {
 
 safeTest('Two instances can run inference simultaneously', {
   timeout: 900_000,
-  skip: isWindowsX64 // TODO: unskip this once we have a new Windows runner with a GPU
+  skip
 }, async t => {
   let addon1 = null
   let addon2 = null
@@ -107,7 +107,7 @@ safeTest('Two instances can run inference simultaneously', {
 
 safeTest('Repeated load/unload cycles should remain stable', {
   timeout: 900_000,
-  skip: isWindowsX64 // TODO: unskip this once we have a new Windows runner with a GPU
+  skip
 }, async t => {
   let currentAddon = null
   try {
@@ -142,7 +142,7 @@ safeTest('Repeated load/unload cycles should remain stable', {
 
 safeTest('Unloading one instance does not affect another generating instance', {
   timeout: 900_000,
-  skip: isWindowsX64 // TODO: unskip this once we have a new Windows runner with a GPU
+  skip
 }, async t => {
   let addon1 = null
   let addon2 = null
@@ -204,7 +204,7 @@ safeTest('Unloading one instance does not affect another generating instance', {
 
 safeTest('Multiple load/unload cycles on one instance while another generates', {
   timeout: 900_000,
-  skip: isWindowsX64 // TODO: unskip this once we have a new Windows runner with a GPU
+  skip
 }, async t => {
   let addon1 = null
   try {
diff --git a/packages/llm-llamacpp/test/integration/qwen3-5-image-elephant-perf.test.js b/packages/llm-llamacpp/test/integration/qwen3-5-image-elephant-perf.test.js
index ade746925c..7c9c898891 100644
--- a/packages/llm-llamacpp/test/integration/qwen3-5-image-elephant-perf.test.js
+++ b/packages/llm-llamacpp/test/integration/qwen3-5-image-elephant-perf.test.js
@@ -4,9 +4,9 @@
 // 30-minute mobile cap. Asserts the elephant keyword + records perf.
 
 const test = require('brittle')
-const { QWEN35_MODEL, IMAGE_CASES, runVlmImagePerf } = require('./_vlm-image-perf.js')
+const { QWEN35_MODEL, IMAGE_CASES, isDarwinX64, runVlmImagePerf } = require('./_vlm-image-perf.js')
 
-test('Qwen3.5-VL image perf [elephant]', { timeout: 1_800_000 }, async t => {
+test('Qwen3.5-VL image perf [elephant]', { timeout: 1_800_000, skip: isDarwinX64 }, async t => {
   await runVlmImagePerf(t, QWEN35_MODEL, IMAGE_CASES.elephant)
 })
 
diff --git a/packages/llm-llamacpp/test/integration/qwen3-5-image-fruit-plate-perf.test.js b/packages/llm-llamacpp/test/integration/qwen3-5-image-fruit-plate-perf.test.js
index 534e9dd681..869953af29 100644
--- a/packages/llm-llamacpp/test/integration/qwen3-5-image-fruit-plate-perf.test.js
+++ b/packages/llm-llamacpp/test/integration/qwen3-5-image-fruit-plate-perf.test.js
@@ -4,9 +4,9 @@
 // the 30-minute mobile cap. Asserts a fruit keyword + records perf.
 
 const test = require('brittle')
-const { QWEN35_MODEL, IMAGE_CASES, runVlmImagePerf } = require('./_vlm-image-perf.js')
+const { QWEN35_MODEL, IMAGE_CASES, isDarwinX64, runVlmImagePerf } = require('./_vlm-image-perf.js')
 
-test('Qwen3.5-VL image perf [fruit plate]', { timeout: 1_800_000 }, async t => {
+test('Qwen3.5-VL image perf [fruit plate]', { timeout: 1_800_000, skip: isDarwinX64 }, async t => {
   await runVlmImagePerf(t, QWEN35_MODEL, IMAGE_CASES['fruit-plate'])
 })
 
diff --git a/packages/llm-llamacpp/test/integration/qwen3-5-image-high-res-aurora-perf.test.js b/packages/llm-llamacpp/test/integration/qwen3-5-image-high-res-aurora-perf.test.js
index 447df2f686..422e38150e 100644
--- a/packages/llm-llamacpp/test/integration/qwen3-5-image-high-res-aurora-perf.test.js
+++ b/packages/llm-llamacpp/test/integration/qwen3-5-image-high-res-aurora-perf.test.js
@@ -4,12 +4,12 @@
 // the 30-minute mobile cap. Asserts an aurora keyword + records perf.
 
 const test = require('brittle')
-const { QWEN35_MODEL, IMAGE_CASES, skipHeavyImages, runVlmImagePerf } = require('./_vlm-image-perf.js')
+const { QWEN35_MODEL, IMAGE_CASES, isDarwinX64, skipHeavyImages, runVlmImagePerf } = require('./_vlm-image-perf.js')
 
 // QVAC-19368: aurora is the heaviest image; skip it on Android on-PR runs
-// where the 30-min Device Farm cap is tight. iOS + desktop always run it.
-// The benchmark (QVAC_PERF_ONLY=true) runs all 3 images on all platforms.
-test('Qwen3.5-VL image perf [high-res aurora]', { timeout: 1_800_000, skip: skipHeavyImages }, async t => {
+// where the 30-min Device Farm cap is tight. Darwin x64 is skipped separately.
+// The benchmark (QVAC_PERF_ONLY=true) runs all 3 images on supported platforms.
+test('Qwen3.5-VL image perf [high-res aurora]', { timeout: 1_800_000, skip: isDarwinX64 || skipHeavyImages }, async t => {
   await runVlmImagePerf(t, QWEN35_MODEL, IMAGE_CASES['high-res-aurora'])
 })
 
diff --git a/packages/llm-llamacpp/test/integration/reasoning.test.js b/packages/llm-llamacpp/test/integration/reasoning.test.js
index 40778cfdad..909cc41204 100644
--- a/packages/llm-llamacpp/test/integration/reasoning.test.js
+++ b/packages/llm-llamacpp/test/integration/reasoning.test.js
@@ -7,7 +7,6 @@ const os = require('bare-os')
 const LlmLlamacpp = require('../../index.js')
 
 const isDarwinX64 = os.platform() === 'darwin' && os.arch() === 'x64'
-const isWindowsX64 = os.platform() === 'win32' && os.arch() === 'x64'
 const isLinuxArm64 = os.platform() === 'linux' && os.arch() === 'arm64'
 const useCpu = isLinuxArm64
 
@@ -126,7 +125,7 @@ function createFollowUpMessages (initialMessages, previousResponse) {
   ]
 }
 safeTest('reasoning tag EOS replacement works with tools=false', {
-  skip: isDarwinX64 || isWindowsX64, // TODO: unskip isWindowsX64 once we have GPU, takes too long
+  skip: isDarwinX64,
   timeout: 600_000
 }, async t => {
   const { inference } = await setupReasoningModel(t, false)
@@ -149,7 +148,7 @@ safeTest('reasoning tag EOS replacement works with tools=false', {
 })
 
 safeTest('reasoning tag EOS replacement works with tools=true', {
-  skip: isDarwinX64 || isWindowsX64, // TODO: unskip isWindowsX64 once we have GPU, takes too long
+  skip: isDarwinX64,
   timeout: 600_000
 }, async t => {
   const { inference } = await setupReasoningModel(t, true)
@@ -172,7 +171,7 @@ safeTest('reasoning tag EOS replacement works with tools=true', {
 })
 
 safeTest('Qwen3 reasoning-budget=0 disables thinking', {
-  skip: isDarwinX64 || isWindowsX64,
+  skip: isDarwinX64,
   timeout: 600_000
 }, async t => {
   const [modelName, dirPath] = await ensureModel({
diff --git a/packages/llm-llamacpp/test/integration/sliding-context.test.js b/packages/llm-llamacpp/test/integration/sliding-context.test.js
index 4e007caa89..76854fedd5 100644
--- a/packages/llm-llamacpp/test/integration/sliding-context.test.js
+++ b/packages/llm-llamacpp/test/integration/sliding-context.test.js
@@ -13,8 +13,7 @@ const isWindowsX64 = platform === 'win32' && arch === 'x64'
 const useCpu = isDarwinX64 || isLinuxArm64
 
 // These are very slow on CI and should be skipped.
-// TODO: unskip Windows once we have a new Windows runner with a GPU
-const skip = isWindowsX64 || isLinuxArm64
+const skip = isDarwinX64 || isLinuxArm64
 
 const DEFAULT_MODEL = {
   name: 'Llama-3.2-1B-Instruct-Q4_0.gguf',
@@ -146,6 +145,7 @@ safeTest('Generation fails with context overflow when sliding disabled', {
   skip
 }, async t => {
   const { model } = await setupModel(t, {
+    ctx_size: '256',
     n_predict: String(SLIDE_PREDICT),
     n_discarded: '0'
   })
diff --git a/packages/llm-llamacpp/test/integration/tools-compact.test.js b/packages/llm-llamacpp/test/integration/tools-compact.test.js
index bd4884462c..dce2826861 100644
--- a/packages/llm-llamacpp/test/integration/tools-compact.test.js
+++ b/packages/llm-llamacpp/test/integration/tools-compact.test.js
@@ -197,6 +197,25 @@ async function runAndCollect (model, prompt, runOptions) {
   }
 }
 
+// Fence for `saveCacheToDisk: true`: the addon resolves `chain.await()` once
+// inference is done, but the session-write hits the disk on a background
+// thread. On a fast runner an immediate `model.unload()` can race the flush
+// and leave a truncated file; the next model's session-load then blocks
+// indefinitely waiting for bytes that never arrive. Poll until size has
+// been stable for one tick to make sure the writer has finalized.
+async function waitForStableSessionFile (filePath, { maxMs = 5000 } = {}) {
+  const start = Date.now()
+  let lastSize = -1
+  while (Date.now() - start < maxMs) {
+    let size = 0
+    try { size = fs.statSync(filePath).size } catch { size = 0 }
+    if (size > 0 && size === lastSize) return size
+    lastSize = size
+    await new Promise(resolve => setTimeout(resolve, 50))
+  }
+  return lastSize
+}
+
 async function runExpectingInvalidPrompt (t, model, prompt, expectedReason, runOptions) {
   cleanupRunOptionsCache(runOptions)
   const response = await model.run(prompt, runOptions)
@@ -759,6 +778,9 @@ safeTest('[tools-compact] session save, destroy, reload with different tools', {
   t.ok(r1.output.length > 0, 'turn 1 produces output')
   t.ok(r1.stats.CacheTokens > 0, 'turn 1 has cache tokens')
 
+  const persistedSize = await waitForStableSessionFile(sessionName)
+  t.ok(persistedSize > 0, `session file flushed to disk (${persistedSize} bytes)`)
+
   await model1.unload()
 
   const { model: model2 } = await setupModel(t)
diff --git a/packages/llm-llamacpp/test/mobile/integration.auto.cjs b/packages/llm-llamacpp/test/mobile/integration.auto.cjs
index 64ec38ab40..886fac0064 100644
--- a/packages/llm-llamacpp/test/mobile/integration.auto.cjs
+++ b/packages/llm-llamacpp/test/mobile/integration.auto.cjs
@@ -381,6 +381,11 @@ async function runConfigParametersTest (options = {}) { // eslint-disable-line n
   return runIntegrationModule('../integration/config-parameters.test.js', options)
 }
 
+async function runContinuousBatchingTest (options = {}) { // eslint-disable-line no-unused-vars
+  if (typeof __shouldRunTest === 'function' && !__shouldRunTest('runContinuousBatchingTest')) return __FILTERED
+  return runIntegrationModule('../integration/continuous-batching.test.js', options)
+}
+
 async function runFinetuningPauseResumeTest (options = {}) { // eslint-disable-line no-unused-vars
   if (typeof __shouldRunTest === 'function' && !__shouldRunTest('runFinetuningPauseResumeTest')) return __FILTERED
   return runIntegrationModule('../integration/finetuning-pause-resume.test.js', options)
diff --git a/packages/llm-llamacpp/test/mobile/test-groups.json b/packages/llm-llamacpp/test/mobile/test-groups.json
index 6696c8bb6a..09c68abc93 100644
--- a/packages/llm-llamacpp/test/mobile/test-groups.json
+++ b/packages/llm-llamacpp/test/mobile/test-groups.json
@@ -7,6 +7,7 @@
     "imageFruitPlate": ["runImageFruitPlateTest"],
     "gemma": ["runGemma4Test"],
     "quantizedKvcache": ["runQuantizedKvcacheTest"],
+    "continuousBatching": ["runContinuousBatchingTest"],
     "lightA": [
       "runApiBehaviorTest",
       "runBitnetTest",
@@ -27,6 +28,7 @@
     ]
   },
   "android": {
+    "continuousBatching": ["runContinuousBatchingTest"],
     "funcShardA": [
       "runFinetuningPauseResumeTest",
       "runConfigParametersTest",
diff --git a/packages/llm-llamacpp/test/unit/CMakeLists.txt b/packages/llm-llamacpp/test/unit/CMakeLists.txt
index 15eafbdcf9..0d04bd7906 100644
--- a/packages/llm-llamacpp/test/unit/CMakeLists.txt
+++ b/packages/llm-llamacpp/test/unit/CMakeLists.txt
@@ -20,6 +20,7 @@ add_executable(
   test_reload_cancel_state.cpp
   test_backend_selection.cpp
   test_tune_config_map.cpp
+  test_generation_params_apply.cpp
   # Placeholder tests (to be implemented)
   test_mtmd_llm_context.cpp
   test_llm_context_base.cpp
@@ -42,6 +43,7 @@ add_executable(
   test_continuous_batching_integration.cpp
   test_kv_leak_on_admit_failure.cpp
   test_runtime_stats.cpp
+  test_reasoning_budget.cpp
   ${CMAKE_SOURCE_DIR}/addon/src/model-interface/AsyncWeightsLoader.cpp
   ${CMAKE_SOURCE_DIR}/addon/src/model-interface/MultiRequestBatcher.cpp
   ${CMAKE_SOURCE_DIR}/addon/src/model-interface/ContinuousBatchScheduler.cpp
diff --git a/packages/llm-llamacpp/test/unit/test_chat_template_utils.cpp b/packages/llm-llamacpp/test/unit/test_chat_template_utils.cpp
index 52a4ebf84f..4c2dda0dbf 100644
--- a/packages/llm-llamacpp/test/unit/test_chat_template_utils.cpp
+++ b/packages/llm-llamacpp/test/unit/test_chat_template_utils.cpp
@@ -220,3 +220,51 @@ TEST_F(
   std::string result = getChatTemplate(nullptr, params, false);
   EXPECT_EQ(result, "my_custom_template");
 }
+
+TEST_F(ChatTemplateUtilsTest, GetPromptExportsQwenThinkingMetadata) {
+  common_chat_templates_ptr tmpls =
+      common_chat_templates_init(nullptr, getFixedQwen3Template());
+  ASSERT_NE(tmpls, nullptr);
+
+  common_chat_templates_inputs inputs;
+  inputs.use_jinja = true;
+  inputs.enable_thinking = true;
+  inputs.add_generation_prompt = true;
+  inputs.messages = {common_chat_msg{
+      /* role = */ "user",
+      /* content = */ "What is the capital of France?",
+  }};
+
+  bool thinkingForcedOpen = true;
+  std::string thinkingStartTag;
+  std::string thinkingEndTag;
+  std::string generationPrompt;
+  const std::string prompt = getPrompt(
+      tmpls.get(),
+      inputs,
+      &thinkingForcedOpen,
+      &thinkingStartTag,
+      &thinkingEndTag,
+      &generationPrompt);
+
+  EXPECT_NE(prompt.find("<|im_start|>assistant"), std::string::npos);
+  EXPECT_EQ(thinkingStartTag, "<think>\n");
+  EXPECT_EQ(thinkingEndTag, "\n</think>\n\n");
+  EXPECT_NE(generationPrompt.find("<|im_start|>assistant"), std::string::npos);
+  EXPECT_FALSE(thinkingForcedOpen);
+}
+
+TEST_F(ChatTemplateUtilsTest, ThinkingForcedOpenTextUsesTemplateSuffix) {
+  EXPECT_EQ(
+      getThinkingForcedOpenText("<|assistant|>\n<reason>\n", "<reason>"),
+      "<reason>\n");
+}
+
+TEST_F(ChatTemplateUtilsTest, ThinkingForcedOpenTextFallsBackToStartTag) {
+  EXPECT_EQ(
+      getThinkingForcedOpenText("<|assistant|>\n", "<reason>"), "<reason>");
+}
+
+TEST_F(ChatTemplateUtilsTest, ThinkingForcedOpenTextEmptyWithoutStartTag) {
+  EXPECT_EQ(getThinkingForcedOpenText("<|assistant|>\n", ""), "");
+}
diff --git a/packages/llm-llamacpp/test/unit/test_generation_params_apply.cpp b/packages/llm-llamacpp/test/unit/test_generation_params_apply.cpp
new file mode 100644
index 0000000000..155e174e1b
--- /dev/null
+++ b/packages/llm-llamacpp/test/unit/test_generation_params_apply.cpp
@@ -0,0 +1,130 @@
+#include <vector>
+
+#include <gtest/gtest.h>
+
+#include "model-interface/GenerationParamsApply.hpp"
+#include "utils/ChatTemplateUtils.hpp"
+
+using qvac_lib_inference_addon_llama::utils::configureReasoningBudgetSampling;
+
+namespace {
+
+std::vector<llama_token> tokens(std::initializer_list<llama_token> values) {
+  return std::vector<llama_token>(values);
+}
+
+} // namespace
+
+TEST(GenerationParamsApplyTest, NoReasoningBudgetOverrideLeavesSamplingState) {
+  common_params_sampling sampling;
+  sampling.reasoning_budget_tokens = 12;
+  sampling.reasoning_budget_start = tokens({1, 2});
+  sampling.reasoning_budget_end = tokens({3});
+  sampling.reasoning_budget_forced = tokens({4, 5});
+  int nPredict = 32;
+
+  GenerationParams overrides;
+  overrides.temp = 0.25f;
+  applyGenerationOverridesToSampling(sampling, nPredict, overrides);
+
+  EXPECT_EQ(sampling.reasoning_budget_tokens, 12);
+  EXPECT_EQ(sampling.reasoning_budget_start, tokens({1, 2}));
+  EXPECT_EQ(sampling.reasoning_budget_end, tokens({3}));
+  EXPECT_EQ(sampling.reasoning_budget_forced, tokens({4, 5}));
+}
+
+TEST(GenerationParamsApplyTest, PositiveReasoningBudgetUpdatesOnlyTokenCap) {
+  common_params_sampling sampling;
+  sampling.reasoning_budget_tokens = -1;
+  sampling.reasoning_budget_start = tokens({10});
+  sampling.reasoning_budget_end = tokens({11});
+  sampling.reasoning_budget_forced = tokens({12});
+  int nPredict = 32;
+
+  GenerationParams overrides;
+  overrides.reasoning_budget = 16;
+  applyGenerationOverridesToSampling(sampling, nPredict, overrides);
+
+  EXPECT_EQ(sampling.reasoning_budget_tokens, 16);
+  EXPECT_EQ(sampling.reasoning_budget_start, tokens({10}));
+  EXPECT_EQ(sampling.reasoning_budget_end, tokens({11}));
+  EXPECT_EQ(sampling.reasoning_budget_forced, tokens({12}));
+}
+
+TEST(GenerationParamsApplyTest, ZeroReasoningBudgetClearsBudgetSamplerState) {
+  common_params_sampling sampling;
+  sampling.reasoning_budget_tokens = 16;
+  sampling.reasoning_budget_start = tokens({10});
+  sampling.reasoning_budget_end = tokens({11});
+  sampling.reasoning_budget_forced = tokens({12});
+  int nPredict = 32;
+
+  GenerationParams overrides;
+  overrides.reasoning_budget = 0;
+  applyGenerationOverridesToSampling(sampling, nPredict, overrides);
+
+  EXPECT_EQ(sampling.reasoning_budget_tokens, -1);
+  EXPECT_TRUE(sampling.reasoning_budget_start.empty());
+  EXPECT_TRUE(sampling.reasoning_budget_end.empty());
+  EXPECT_TRUE(sampling.reasoning_budget_forced.empty());
+}
+
+TEST(
+    GenerationParamsApplyTest,
+    UnrestrictedReasoningBudgetClearsBudgetSamplerState) {
+  common_params_sampling sampling;
+  sampling.reasoning_budget_tokens = 16;
+  sampling.reasoning_budget_start = tokens({10});
+  sampling.reasoning_budget_end = tokens({11});
+  sampling.reasoning_budget_forced = tokens({12});
+  int nPredict = 32;
+
+  GenerationParams overrides;
+  overrides.reasoning_budget = -1;
+  applyGenerationOverridesToSampling(sampling, nPredict, overrides);
+
+  EXPECT_EQ(sampling.reasoning_budget_tokens, -1);
+  EXPECT_TRUE(sampling.reasoning_budget_start.empty());
+  EXPECT_TRUE(sampling.reasoning_budget_end.empty());
+  EXPECT_TRUE(sampling.reasoning_budget_forced.empty());
+}
+
+TEST(
+    GenerationParamsApplyTest,
+    ConfigureReasoningBudgetSamplingClearsStaleStateWhenDisabled) {
+  common_params params;
+  params.reasoning_budget = 0;
+  params.sampling.reasoning_budget_tokens = 16;
+  params.sampling.reasoning_budget_start = tokens({10});
+  params.sampling.reasoning_budget_end = tokens({11});
+  params.sampling.reasoning_budget_forced = tokens({12});
+  params.sampling.generation_prompt = "<assistant><think>";
+
+  EXPECT_TRUE(configureReasoningBudgetSampling(
+      params, nullptr, "<think>", "</think>", "<assistant><think>"));
+  EXPECT_EQ(params.sampling.reasoning_budget_tokens, -1);
+  EXPECT_TRUE(params.sampling.reasoning_budget_start.empty());
+  EXPECT_TRUE(params.sampling.reasoning_budget_end.empty());
+  EXPECT_TRUE(params.sampling.reasoning_budget_forced.empty());
+  EXPECT_TRUE(params.sampling.generation_prompt.empty());
+}
+
+TEST(
+    GenerationParamsApplyTest,
+    ConfigureReasoningBudgetSamplingKeepsPositiveCapWithoutContext) {
+  common_params params;
+  params.reasoning_budget = 8;
+  params.sampling.reasoning_budget_tokens = -1;
+  params.sampling.reasoning_budget_start = tokens({10});
+  params.sampling.reasoning_budget_end = tokens({11});
+  params.sampling.reasoning_budget_forced = tokens({12});
+  params.sampling.generation_prompt = "<assistant><think>";
+
+  EXPECT_TRUE(configureReasoningBudgetSampling(
+      params, nullptr, "<think>", "</think>", "<assistant><think>"));
+  EXPECT_EQ(params.sampling.reasoning_budget_tokens, 8);
+  EXPECT_TRUE(params.sampling.reasoning_budget_start.empty());
+  EXPECT_TRUE(params.sampling.reasoning_budget_end.empty());
+  EXPECT_TRUE(params.sampling.reasoning_budget_forced.empty());
+  EXPECT_TRUE(params.sampling.generation_prompt.empty());
+}
diff --git a/packages/llm-llamacpp/test/unit/test_reasoning_budget.cpp b/packages/llm-llamacpp/test/unit/test_reasoning_budget.cpp
new file mode 100644
index 0000000000..2e94a9ea5f
--- /dev/null
+++ b/packages/llm-llamacpp/test/unit/test_reasoning_budget.cpp
@@ -0,0 +1,141 @@
+#include <filesystem>
+#include <memory>
+#include <stdexcept>
+#include <string>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+
+#include <gtest/gtest.h>
+
+#include "common/common.h"
+#include "model-interface/LlamaModel.hpp"
+#include "test_common.hpp"
+
+namespace fs = std::filesystem;
+
+namespace {
+
+constexpr int REASONING_BUDGET = 3;
+constexpr int PREDICT_TOKENS = 10;
+constexpr const char* THINKING_START_TAG = "<think>\n";
+constexpr const char* THINKING_END_TAG = "\n</think>\n\n";
+
+std::string sliceReasoning(const std::string& text) {
+  const size_t open = text.find(THINKING_START_TAG);
+  if (open == std::string::npos) {
+    return "";
+  }
+
+  const size_t bodyStart = open + std::string(THINKING_START_TAG).size();
+  const size_t close = text.find(THINKING_END_TAG, bodyStart);
+  if (close == std::string::npos || close < bodyStart) {
+    return "";
+  }
+
+  return text.substr(bodyStart, close - bodyStart);
+}
+
+size_t countTokens(llama_context* ctx, const std::string& text) {
+  return common_tokenize(ctx, text, false, true).size();
+}
+
+struct RunResult {
+  std::string output;
+  size_t reasoningTokens = 0;
+};
+
+} // namespace
+
+class ReasoningBudgetModelTest : public ::testing::Test {
+protected:
+  void SetUp() override {
+    using MP = test_common::TestModelPath;
+    qwen3Model_ =
+        MP("Qwen3-0.6B-Q8_0.gguf",
+           "QWEN3_MODEL_PATH",
+           MP::OnMissing::Skip,
+           "https://huggingface.co/Qwen/Qwen3-0.6B-GGUF");
+
+    config_["ctx_size"] = "4096";
+    config_["n_predict"] = std::to_string(PREDICT_TOKENS);
+    config_["seed"] = "50";
+    config_["temp"] = "0";
+    config_["top_p"] = "1";
+    config_["device"] = test_common::getTestDevice();
+    config_["gpu_layers"] = test_common::getTestGpuLayers();
+    config_["backendsDir"] = test_common::getTestBackendsDir().string();
+  }
+
+  [[nodiscard]] bool hasQwen3Model() const {
+    return qwen3Model_.found() && fs::exists(qwen3Model_.path);
+  }
+
+  std::unique_ptr<LlamaModel>
+  createModel(std::unordered_map<std::string, std::string> config) {
+    auto model = std::make_unique<LlamaModel>(
+        std::string(qwen3Model_.path), std::string(), std::move(config));
+    model->waitForLoadInitialization();
+    return model;
+  }
+
+  RunResult runBudgetedPrompt(
+      std::unordered_map<std::string, std::string> config,
+      const GenerationParams& generationParams = {}) {
+    auto model = createModel(std::move(config));
+    if (!model->isLoaded()) {
+      throw std::runtime_error("Qwen3 model failed to load");
+    }
+
+    LlamaModel::Prompt prompt;
+    prompt.input =
+        R"([{"role":"system","content":"You are a helpful assistant."},{"role":"user","content":"What is the capital of France? Answer in one word."}])";
+    prompt.generationParams = generationParams;
+
+    const std::string output = model->processPrompt(prompt);
+    const std::string reasoning = sliceReasoning(output);
+    return {output, countTokens(model->getContext(), reasoning)};
+  }
+
+  test_common::TestModelPath qwen3Model_;
+  std::unordered_map<std::string, std::string> config_;
+};
+
+TEST_F(
+    ReasoningBudgetModelTest,
+    LoadTimeBudgetThreeGeneratesThreeReasoningTokens) {
+  if (!hasQwen3Model()) {
+    GTEST_SKIP() << qwen3Model_.missingMessage();
+  }
+
+  auto config = config_;
+  config["reasoning-budget"] = std::to_string(REASONING_BUDGET);
+
+  const RunResult result = runBudgetedPrompt(std::move(config));
+
+  EXPECT_NE(result.output.find(THINKING_START_TAG), std::string::npos)
+      << result.output;
+  EXPECT_NE(result.output.find(THINKING_END_TAG), std::string::npos)
+      << result.output;
+  EXPECT_EQ(result.reasoningTokens, REASONING_BUDGET) << result.output;
+}
+
+TEST_F(
+    ReasoningBudgetModelTest,
+    PerRequestBudgetThreeGeneratesThreeReasoningTokens) {
+  if (!hasQwen3Model()) {
+    GTEST_SKIP() << qwen3Model_.missingMessage();
+  }
+
+  GenerationParams generationParams;
+  generationParams.reasoning_budget = REASONING_BUDGET;
+  generationParams.n_predict = PREDICT_TOKENS;
+
+  const RunResult result = runBudgetedPrompt(config_, generationParams);
+
+  EXPECT_NE(result.output.find(THINKING_START_TAG), std::string::npos)
+      << result.output;
+  EXPECT_NE(result.output.find(THINKING_END_TAG), std::string::npos)
+      << result.output;
+  EXPECT_EQ(result.reasoningTokens, REASONING_BUDGET) << result.output;
+}
diff --git a/packages/llm-llamacpp/test/unit/test_tune_config_map.cpp b/packages/llm-llamacpp/test/unit/test_tune_config_map.cpp
index 5f194d88e3..8262920a13 100644
--- a/packages/llm-llamacpp/test/unit/test_tune_config_map.cpp
+++ b/packages/llm-llamacpp/test/unit/test_tune_config_map.cpp
@@ -3,6 +3,7 @@
 #include <unordered_map>
 
 #include <gtest/gtest.h>
+#include <inference-addon-cpp/Errors.hpp>
 
 #include "model-interface/LlamaModel.hpp"
 #include "test_common.hpp"
@@ -15,32 +16,33 @@ class TuneConfigMapTest : public ::testing::Test {
   std::unordered_map<std::string, std::string> configFilemap_;
 };
 
-// ---- Non-BitNet: no modifications ----
-
-TEST_F(TuneConfigMapTest, NonBitnet_NoChanges) {
+TEST_F(TuneConfigMapTest, NonBitnet_FlashAttnDefaultsOn) {
   MockModelMetaData meta(false, "llama");
 
   LlamaModel::tuneConfigMap(configFilemap_, meta, std::nullopt);
 
-  EXPECT_EQ(configFilemap_.count("flash-attn"), 0);
+  ASSERT_EQ(configFilemap_.count("flash-attn"), 1);
+  EXPECT_EQ(configFilemap_["flash-attn"], "on");
   EXPECT_EQ(configFilemap_.count("ubatch-size"), 0);
 }
 
-TEST_F(TuneConfigMapTest, OneBitButNotBitnetArch_NoChanges) {
+TEST_F(TuneConfigMapTest, OneBitButNotBitnetArch_FlashAttnDefaultsOn) {
   MockModelMetaData meta(true, "llama");
 
   LlamaModel::tuneConfigMap(configFilemap_, meta, 830);
 
-  EXPECT_EQ(configFilemap_.count("flash-attn"), 0);
+  ASSERT_EQ(configFilemap_.count("flash-attn"), 1);
+  EXPECT_EQ(configFilemap_["flash-attn"], "on");
   EXPECT_EQ(configFilemap_.count("ubatch-size"), 0);
 }
 
-TEST_F(TuneConfigMapTest, BitnetArchButNotOneBit_NoChanges) {
+TEST_F(TuneConfigMapTest, BitnetArchButNotOneBit_FlashAttnDefaultsOn) {
   MockModelMetaData meta(false, "bitnet");
 
   LlamaModel::tuneConfigMap(configFilemap_, meta, 830);
 
-  EXPECT_EQ(configFilemap_.count("flash-attn"), 0);
+  ASSERT_EQ(configFilemap_.count("flash-attn"), 1);
+  EXPECT_EQ(configFilemap_["flash-attn"], "on");
   EXPECT_EQ(configFilemap_.count("ubatch-size"), 0);
 }
 
@@ -190,16 +192,16 @@ TEST_F(TuneConfigMapTest, Bitnet_Adreno799_UbatchUnchanged) {
   EXPECT_EQ(configFilemap_.count("ubatch-size"), 0);
 }
 
-// ---- OpenCL backend: flash-attn disabled by default unless user-set ----
+// ---- OpenCL backend: flash-attn defaulted ON like every other GPU path ----
 
-TEST_F(TuneConfigMapTest, OpenCl_NonBitnet_FlashAttnDisabledByDefault) {
+TEST_F(TuneConfigMapTest, OpenCl_NonBitnet_FlashAttnDefaultsOn) {
   MockModelMetaData meta(false, "llama");
 
   LlamaModel::tuneConfigMap(
       configFilemap_, meta, std::nullopt, FtOverrides{}, /*isOpenCl=*/true);
 
   ASSERT_EQ(configFilemap_.count("flash-attn"), 1);
-  EXPECT_EQ(configFilemap_["flash-attn"], "off");
+  EXPECT_EQ(configFilemap_["flash-attn"], "on");
 }
 
 TEST_F(TuneConfigMapTest, OpenCl_UserSetFlashAttnHyphen_Respected) {
@@ -223,13 +225,107 @@ TEST_F(TuneConfigMapTest, OpenCl_UserSetFlashAttnUnderscore_Respected) {
   EXPECT_EQ(configFilemap_["flash_attn"], "on");
 }
 
-TEST_F(TuneConfigMapTest, NotOpenCl_NonBitnet_FlashAttnUnchanged) {
+TEST_F(TuneConfigMapTest, NotOpenCl_NonBitnet_FlashAttnDefaultsOn) {
   MockModelMetaData meta(false, "llama");
 
   LlamaModel::tuneConfigMap(
       configFilemap_, meta, std::nullopt, FtOverrides{}, /*isOpenCl=*/false);
 
-  EXPECT_EQ(configFilemap_.count("flash-attn"), 0);
+  ASSERT_EQ(configFilemap_.count("flash-attn"), 1);
+  EXPECT_EQ(configFilemap_["flash-attn"], "on");
+}
+
+TEST_F(TuneConfigMapTest, OpenCl_RejectsStandardQuantizedKCache) {
+  MockModelMetaData meta(false, "llama");
+  configFilemap_["cache-type-k"] = "q8_0";
+
+  EXPECT_THROW(
+      LlamaModel::tuneConfigMap(
+          configFilemap_, meta, std::nullopt, FtOverrides{}, /*isOpenCl=*/true),
+      qvac_errors::StatusError);
+}
+
+TEST_F(TuneConfigMapTest, OpenCl_RejectsStandardQuantizedVCacheUnderscore) {
+  MockModelMetaData meta(false, "llama");
+  configFilemap_["cache_type_v"] = "q4_0";
+
+  EXPECT_THROW(
+      LlamaModel::tuneConfigMap(
+          configFilemap_, meta, std::nullopt, FtOverrides{}, /*isOpenCl=*/true),
+      qvac_errors::StatusError);
+}
+
+TEST_F(TuneConfigMapTest, OpenCl_AllowsNonQuantizedCacheTypes) {
+  MockModelMetaData meta(false, "llama");
+  configFilemap_["cache-type-k"] = "f16";
+  configFilemap_["cache-type-v"] = "bf16";
+
+  EXPECT_NO_THROW(
+      LlamaModel::tuneConfigMap(
+          configFilemap_,
+          meta,
+          std::nullopt,
+          FtOverrides{},
+          /*isOpenCl=*/true));
+}
+
+TEST_F(TuneConfigMapTest, NotOpenCl_AllowsStandardQuantizedCacheTypes) {
+  MockModelMetaData meta(false, "llama");
+  configFilemap_["cache-type-k"] = "q8_0";
+  configFilemap_["cache-type-v"] = "q4_0";
+
+  EXPECT_NO_THROW(
+      LlamaModel::tuneConfigMap(
+          configFilemap_,
+          meta,
+          std::nullopt,
+          FtOverrides{},
+          /*isOpenCl=*/false));
+}
+
+TEST_F(TuneConfigMapTest, NotOpenClNotMetal_AllowsTurboQuantCacheTypes) {
+  MockModelMetaData meta(false, "llama");
+  configFilemap_["cache-type-k"] = "tbq4_0";
+  configFilemap_["cache-type-v"] = "pq4_0";
+
+  EXPECT_NO_THROW(
+      LlamaModel::tuneConfigMap(
+          configFilemap_,
+          meta,
+          std::nullopt,
+          FtOverrides{},
+          /*isOpenCl=*/false,
+          /*isMetal=*/false));
+}
+
+TEST_F(TuneConfigMapTest, Metal_RejectsTurboQuantCacheTypes) {
+  MockModelMetaData meta(false, "llama");
+  configFilemap_["cache-type-k"] = "tbq4_0";
+
+  EXPECT_THROW(
+      LlamaModel::tuneConfigMap(
+          configFilemap_,
+          meta,
+          std::nullopt,
+          FtOverrides{},
+          /*isOpenCl=*/false,
+          /*isMetal=*/true),
+      qvac_errors::StatusError);
+}
+
+TEST_F(TuneConfigMapTest, Metal_AllowsStandardQuantizedCacheTypes) {
+  MockModelMetaData meta(false, "llama");
+  configFilemap_["cache-type-k"] = "q8_0";
+  configFilemap_["cache-type-v"] = "q4_0";
+
+  EXPECT_NO_THROW(
+      LlamaModel::tuneConfigMap(
+          configFilemap_,
+          meta,
+          std::nullopt,
+          FtOverrides{},
+          /*isOpenCl=*/false,
+          /*isMetal=*/true));
 }
 
 // ---- Finetuning: flash-attn disabled for any architecture ----