diff --git a/lua/claude_handler.lua b/lua/claude_handler.lua
index 8531a1c..4f144ad 100644
--- a/lua/claude_handler.lua
+++ b/lua/claude_handler.lua
@@ -49,6 +49,10 @@ function _M.handle()
 
     -- Translate Claude request to OpenAI format
     local oai_request = claude_fmt.request_to_openai(claude_body)
+
+    -- Thinking mode handling (operates on OpenAI request format)
+    oai_request, model = e2ee.handle_thinking(oai_request)
+
     local oai_body = cjson.encode(oai_request)
 
     if not is_streaming then
diff --git a/lua/e2ee_discovery.lua b/lua/e2ee_discovery.lua
index 8ee8d71..832fb03 100644
--- a/lua/e2ee_discovery.lua
+++ b/lua/e2ee_discovery.lua
@@ -99,11 +99,18 @@ function _M.resolve_chute_id(model, api_key)
         return model
     end
 
+    -- Strip the :THINKING suffix before map lookup. v1/models returns
+    -- base model IDs only; :THINKING is a proxy-side flag and must not
+    -- leak into the lookup key. Other suffixes (e.g. LoRA names) are
+    -- left intact so they surface as proper errors upstream rather than
+    -- silently resolving to the base chute.
+    local base_model = model:match("^(.-):THINKING$") or model
+
     local now = ngx.now()
 
     -- Check cache
     if model_map and now < model_map_expires then
-        local entry = model_map[model]
+        local entry = model_map[base_model]
         if entry then
             return check_confidential(model, entry)
         end
@@ -114,7 +121,7 @@ function _M.resolve_chute_id(model, api_key)
     if not map then
         -- If we have a stale cache, try it
         if model_map then
-            local entry = model_map[model]
+            local entry = model_map[base_model]
             if entry then
                 return check_confidential(model, entry)
             end
@@ -125,7 +132,7 @@ function _M.resolve_chute_id(model, api_key)
     model_map = map
     model_map_expires = now + MODEL_MAP_TTL
 
-    local entry = map[model]
+    local entry = map[base_model]
     return check_confidential(model, entry)
 end
 
diff --git a/lua/e2ee_handler.lua b/lua/e2ee_handler.lua
index 71938e5..b605c9f 100644
--- a/lua/e2ee_handler.lua
+++ b/lua/e2ee_handler.lua
@@ -17,6 +17,31 @@ local API_BASE = "https://api.chutes.ai"
 
 local _M = {}
 
+-- Per-model thinking defaults. Mirrors chutes-api's invocation router
+-- (api/invocation/router.py). The /v1/models endpoint exposes a
+-- "reasoning" capability flag in supported_features but not the
+-- per-model default value, so these stay duplicated until the API
+-- grows a dedicated attribute.
+local THINKING_DEFAULT_ON_PREFIXES = {
+    "deepseek-ai/DeepSeek-V3.2-Speciale",
+    "zai-org/GLM-4.7",
+    "moonshotai/Kimi-K2.5",
+}
+local THINKING_DEFAULT_ON_EXACT = {
+    ["deepseek-ai/DeepSeek-V3.2-Speciale"] = true,
+    ["deepseek-ai/DeepSeek-V3.2-Speciale-TEE"] = true,
+    ["zai-org/GLM-4.7"] = true,
+    ["zai-org/GLM-4.7-TEE"] = true,
+    ["moonshotai/Kimi-K2.5"] = true,
+    ["moonshotai/Kimi-K2.5-TEE"] = true,
+}
+local MIMO_PREFIX = "XiaomiMiMo/MiMo-V2-Flash"
+local MIMO_EXACT_TEE = "XiaomiMiMo/MiMo-V2-Flash-TEE"
+
+local function starts_with(s, prefix)
+    return s:sub(1, #prefix) == prefix
+end
+
 --- Extract API key from Authorization header or x-api-key
 function _M.get_api_key()
     local headers = ngx.req.get_headers()
@@ -40,6 +65,68 @@ function _M.get_api_key()
     return key
 end
 
+--- Handle thinking mode for OpenAI format requests
+-- Takes an OpenAI request table, applies thinking logic, and returns the modified table
+-- Also strips :THINKING suffix from model name and returns the model name
+function _M.handle_thinking(oai_request)
+    if not oai_request or not oai_request.model then
+        return oai_request, nil
+    end
+
+    local model = oai_request.model
+    local enable_thinking = false
+    local think_header = ngx.req.get_headers()["X-Enable-Thinking"]
+    if think_header and think_header:lower() == "true" then
+        enable_thinking = true
+    end
+    if model:match(":THINKING$") then
+        model = model:sub(1, -#":THINKING" - 1)
+        oai_request.model = model
+        enable_thinking = true
+    end
+    if enable_thinking then
+        if not oai_request.chat_template_kwargs then
+            oai_request.chat_template_kwargs = {}
+        end
+        oai_request.chat_template_kwargs.thinking = true
+        oai_request.chat_template_kwargs.enable_thinking = true
+    end
+
+    local kwargs = oai_request.chat_template_kwargs
+    if kwargs then
+        -- Normalize the two spellings so downstream sees both.
+        if kwargs.thinking ~= nil and kwargs.enable_thinking == nil then
+            kwargs.enable_thinking = kwargs.thinking
+        end
+        if kwargs.enable_thinking ~= nil and kwargs.thinking == nil then
+            kwargs.thinking = kwargs.enable_thinking
+        end
+
+        if kwargs.thinking == nil then
+            for _, prefix in ipairs(THINKING_DEFAULT_ON_PREFIXES) do
+                if starts_with(model, prefix) then
+                    kwargs.thinking = true
+                    kwargs.enable_thinking = true
+                    break
+                end
+            end
+        end
+
+        if kwargs.thinking == nil and starts_with(model, MIMO_PREFIX) then
+            kwargs.thinking = false
+            kwargs.enable_thinking = false
+        end
+    else
+        if THINKING_DEFAULT_ON_EXACT[model] then
+            oai_request.chat_template_kwargs = { thinking = true, enable_thinking = true }
+        elseif model == MIMO_EXACT_TEE then
+            oai_request.chat_template_kwargs = { thinking = false, enable_thinking = false }
+        end
+    end
+
+    return oai_request, model
+end
+
 --- Send error response
 function _M.send_error(status, message)
     ngx.status = status
@@ -345,6 +432,12 @@ function _M.handle()
         return _M.send_error(400, "missing 'model' field")
     end
 
+    -- Thinking mode handling
+    payload, model = _M.handle_thinking(payload)
+
+    -- Re-encode body with any chat_template_kwargs modifications
+    body = cjson.encode(payload)
+
     local is_streaming = (payload.stream == true)
     local original_path = ngx.var.uri
 
diff --git a/lua/responses_handler.lua b/lua/responses_handler.lua
index 10bf166..02d3a71 100644
--- a/lua/responses_handler.lua
+++ b/lua/responses_handler.lua
@@ -49,6 +49,10 @@ function _M.handle()
 
     -- Translate Responses request to OpenAI format
     local oai_request = resp_fmt.request_to_openai(resp_body)
+
+    -- Thinking mode handling (operates on OpenAI request format)
+    oai_request, model = e2ee.handle_thinking(oai_request)
+
     local oai_body = cjson.encode(oai_request)
 
     if not is_streaming then
diff --git a/tests/test_thinking.lua b/tests/test_thinking.lua
new file mode 100644
index 0000000..04d6dc8
--- /dev/null
+++ b/tests/test_thinking.lua
@@ -0,0 +1,271 @@
+--
+-- Unit tests for e2ee_handler.handle_thinking
+--
+-- Runs the :THINKING suffix / X-Enable-Thinking header logic in isolation
+-- by stubbing the openresty globals (ngx, cjson.safe, resty.http) and the
+-- sibling lua modules that e2ee_handler requires but doesn't exercise here.
+--
+-- Usage:
+--   luajit tests/test_thinking.lua
+--
+-- Exits non-zero on any assertion failure.
+--
+
+-- Make lua/ importable regardless of cwd.
+local script_path = arg and arg[0] or ""
+local script_dir = script_path:match("(.*/)") or "./"
+package.path = script_dir .. "../lua/?.lua;" .. package.path
+
+-- Stub ngx. The only surface handle_thinking touches is ngx.req.get_headers().
+local stub_headers = {}
+_G.ngx = {
+    req = {
+        get_headers = function() return stub_headers end,
+    },
+    log = function() end,
+    INFO = 0, WARN = 0, ERR = 0,
+    now = function() return 0 end,
+}
+
+-- Stub modules that e2ee_handler requires but handle_thinking doesn't use.
+package.loaded["cjson.safe"] = { encode = function(x) return x end, decode = function(x) return x end }
+package.loaded["resty.http"] = { new = function() return {} end }
+package.loaded["e2ee_crypto"] = {}
+package.loaded["e2ee_discovery"] = {}
+
+local e2ee = require("e2ee_handler")
+
+local failures = 0
+local total = 0
+
+local function set_header(name, value)
+    stub_headers = {}
+    if value ~= nil then stub_headers[name] = value end
+end
+
+local function assert_eq(label, actual, expected)
+    total = total + 1
+    local ok
+    if type(expected) == "table" and type(actual) == "table" then
+        ok = true
+        for k, v in pairs(expected) do
+            if actual[k] ~= v then ok = false; break end
+        end
+        if ok then
+            for k, v in pairs(actual) do
+                if expected[k] ~= v then ok = false; break end
+            end
+        end
+    else
+        ok = (actual == expected)
+    end
+    if not ok then
+        failures = failures + 1
+        io.stderr:write(string.format(
+            "FAIL %s\n  expected: %s\n  actual:   %s\n",
+            label, tostring(expected), tostring(actual)
+        ))
+    else
+        io.stdout:write("ok   " .. label .. "\n")
+    end
+end
+
+-- 1. Missing model → returns request, nil; no mutation.
+set_header()
+do
+    local req = { stream = true }
+    local out, model = e2ee.handle_thinking(req)
+    assert_eq("nil model returns nil", model, nil)
+    assert_eq("nil model leaves request untouched", out.chat_template_kwargs, nil)
+end
+
+-- 2. Plain model, no suffix, no header, no kwargs → unchanged.
+set_header()
+do
+    local req = { model = "deepseek-ai/DeepSeek-V3.1-TEE" }
+    local out, model = e2ee.handle_thinking(req)
+    assert_eq("plain model name preserved", model, "deepseek-ai/DeepSeek-V3.1-TEE")
+    assert_eq("plain model payload.model preserved", out.model, "deepseek-ai/DeepSeek-V3.1-TEE")
+    assert_eq("plain model leaves kwargs absent", out.chat_template_kwargs, nil)
+end
+
+-- 3. :THINKING suffix → strip, set both keys.
+set_header()
+do
+    local req = { model = "zai-org/GLM-5.1-TEE:THINKING" }
+    local out, model = e2ee.handle_thinking(req)
+    assert_eq(":THINKING strips suffix (return)", model, "zai-org/GLM-5.1-TEE")
+    assert_eq(":THINKING strips suffix (payload)", out.model, "zai-org/GLM-5.1-TEE")
+    assert_eq(":THINKING sets thinking=true", out.chat_template_kwargs.thinking, true)
+    assert_eq(":THINKING sets enable_thinking=true", out.chat_template_kwargs.enable_thinking, true)
+end
+
+-- 4. X-Enable-Thinking: true → sets both keys, no suffix.
+set_header("X-Enable-Thinking", "true")
+do
+    local req = { model = "zai-org/GLM-5.1-TEE" }
+    local out = e2ee.handle_thinking(req)
+    assert_eq("header=true sets thinking", out.chat_template_kwargs.thinking, true)
+    assert_eq("header=true sets enable_thinking", out.chat_template_kwargs.enable_thinking, true)
+    assert_eq("header=true keeps model intact", out.model, "zai-org/GLM-5.1-TEE")
+end
+
+-- 5. X-Enable-Thinking: TRUE (uppercase) → case-insensitive.
+set_header("X-Enable-Thinking", "TRUE")
+do
+    local req = { model = "zai-org/GLM-5.1-TEE" }
+    local out = e2ee.handle_thinking(req)
+    assert_eq("header TRUE case-insensitive", out.chat_template_kwargs.thinking, true)
+end
+
+-- 6. X-Enable-Thinking: false → does not set kwargs.
+set_header("X-Enable-Thinking", "false")
+do
+    local req = { model = "deepseek-ai/DeepSeek-V3.1-TEE" }
+    local out = e2ee.handle_thinking(req)
+    assert_eq("header=false leaves kwargs absent", out.chat_template_kwargs, nil)
+end
+
+-- 7. Only `thinking` key → `enable_thinking` propagated.
+set_header()
+do
+    local req = {
+        model = "deepseek-ai/DeepSeek-V3.1-TEE",
+        chat_template_kwargs = { thinking = true },
+    }
+    local out = e2ee.handle_thinking(req)
+    assert_eq("thinking=true propagates to enable_thinking", out.chat_template_kwargs.enable_thinking, true)
+end
+
+-- 8. Only `enable_thinking` key → `thinking` propagated.
+set_header()
+do
+    local req = {
+        model = "deepseek-ai/DeepSeek-V3.1-TEE",
+        chat_template_kwargs = { enable_thinking = false },
+    }
+    local out = e2ee.handle_thinking(req)
+    assert_eq("enable_thinking=false propagates to thinking", out.chat_template_kwargs.thinking, false)
+end
+
+-- 9. Per-model prefix default-on (GLM-4.7) when kwargs present but no thinking key.
+set_header()
+do
+    local req = {
+        model = "zai-org/GLM-4.7",
+        chat_template_kwargs = { some_other = "x" },
+    }
+    local out = e2ee.handle_thinking(req)
+    assert_eq("GLM-4.7 prefix defaults thinking=true", out.chat_template_kwargs.thinking, true)
+    assert_eq("GLM-4.7 prefix defaults enable_thinking=true", out.chat_template_kwargs.enable_thinking, true)
+    assert_eq("GLM-4.7 prefix preserves other keys", out.chat_template_kwargs.some_other, "x")
+end
+
+-- 10. Per-model exact default-on (GLM-4.7-TEE) when kwargs absent.
+set_header()
+do
+    local req = { model = "zai-org/GLM-4.7-TEE" }
+    local out = e2ee.handle_thinking(req)
+    assert_eq("GLM-4.7-TEE exact defaults thinking=true", out.chat_template_kwargs.thinking, true)
+    assert_eq("GLM-4.7-TEE exact defaults enable_thinking=true", out.chat_template_kwargs.enable_thinking, true)
+end
+
+-- 11. Per-model exact default-on for DeepSeek-V3.2-Speciale-TEE and Kimi-K2.5-TEE.
+set_header()
+do
+    local req = { model = "deepseek-ai/DeepSeek-V3.2-Speciale-TEE" }
+    local out = e2ee.handle_thinking(req)
+    assert_eq("DeepSeek-V3.2-Speciale-TEE exact thinking=true", out.chat_template_kwargs.thinking, true)
+end
+do
+    local req = { model = "moonshotai/Kimi-K2.5-TEE" }
+    local out = e2ee.handle_thinking(req)
+    assert_eq("Kimi-K2.5-TEE exact thinking=true", out.chat_template_kwargs.thinking, true)
+end
+
+-- 12. MiMo-V2-Flash prefix default-off when kwargs present.
+set_header()
+do
+    local req = {
+        model = "XiaomiMiMo/MiMo-V2-Flash-Pro",
+        chat_template_kwargs = { some_other = "x" },
+    }
+    local out = e2ee.handle_thinking(req)
+    assert_eq("MiMo-V2-Flash prefix defaults thinking=false", out.chat_template_kwargs.thinking, false)
+    assert_eq("MiMo-V2-Flash prefix defaults enable_thinking=false", out.chat_template_kwargs.enable_thinking, false)
+end
+
+-- 13. MiMo-V2-Flash-TEE exact default-off when kwargs absent.
+set_header()
+do
+    local req = { model = "XiaomiMiMo/MiMo-V2-Flash-TEE" }
+    local out = e2ee.handle_thinking(req)
+    assert_eq("MiMo-V2-Flash-TEE exact thinking=false", out.chat_template_kwargs.thinking, false)
+    assert_eq("MiMo-V2-Flash-TEE exact enable_thinking=false", out.chat_template_kwargs.enable_thinking, false)
+end
+
+-- 14. :THINKING on MiMo → user override wins over default-off.
+set_header()
+do
+    local req = { model = "XiaomiMiMo/MiMo-V2-Flash-TEE:THINKING" }
+    local out, model = e2ee.handle_thinking(req)
+    assert_eq(":THINKING on MiMo strips suffix", model, "XiaomiMiMo/MiMo-V2-Flash-TEE")
+    assert_eq(":THINKING on MiMo sets thinking=true", out.chat_template_kwargs.thinking, true)
+end
+
+-- 15. Non-reasoning model + no suffix + no header → no kwargs injected.
+set_header()
+do
+    local req = { model = "deepseek-ai/DeepSeek-V3.1-TEE" }
+    local out = e2ee.handle_thinking(req)
+    assert_eq("non-reasoning model leaves kwargs absent", out.chat_template_kwargs, nil)
+end
+
+-- 16. :THINKING preserves unrelated kwargs.
+set_header()
+do
+    local req = {
+        model = "zai-org/GLM-5.1-TEE:THINKING",
+        chat_template_kwargs = { tools_in_user_message = false },
+    }
+    local out = e2ee.handle_thinking(req)
+    assert_eq(":THINKING preserves unrelated kwargs",
+        out.chat_template_kwargs.tools_in_user_message, false)
+    assert_eq(":THINKING still sets thinking=true",
+        out.chat_template_kwargs.thinking, true)
+end
+
+-- 17. :THINKING appears mid-string (not suffix) → not stripped.
+set_header()
+do
+    local req = { model = "some:THINKING:other" }
+    local out, model = e2ee.handle_thinking(req)
+    assert_eq("non-suffix :THINKING not stripped", model, "some:THINKING:other")
+    assert_eq("non-suffix :THINKING leaves kwargs absent", out.chat_template_kwargs, nil)
+end
+
+-- 18. Explicit thinking=false overrides per-model default-on.
+set_header()
+do
+    local req = {
+        model = "zai-org/GLM-4.7",
+        chat_template_kwargs = { thinking = false },
+    }
+    local out = e2ee.handle_thinking(req)
+    assert_eq("explicit thinking=false overrides GLM-4.7 default", out.chat_template_kwargs.thinking, false)
+    assert_eq("explicit thinking=false propagates enable_thinking=false", out.chat_template_kwargs.enable_thinking, false)
+end
+
+-- 19. DeepSeek-V3.2-Speciale prefix (bare, no TEE) with kwargs present.
+set_header()
+do
+    local req = {
+        model = "deepseek-ai/DeepSeek-V3.2-Speciale",
+        chat_template_kwargs = {},
+    }
+    local out = e2ee.handle_thinking(req)
+    assert_eq("DeepSeek-V3.2-Speciale prefix default thinking=true", out.chat_template_kwargs.thinking, true)
+end
+
+io.stdout:write(string.format("\n%d/%d passed\n", total - failures, total))
+os.exit(failures == 0 and 0 or 1)