diff --git a/workloads/devstral_small_2_24b_b200.yaml b/workloads/devstral_small_2_24b_b200.yaml
new file mode 100644
index 0000000..de7ac8b
--- /dev/null
+++ b/workloads/devstral_small_2_24b_b200.yaml
@@ -0,0 +1,29 @@
+# Devstral-Small-2-24B-Instruct-2512 (FP8) on B200 (1xB200, TP=1) -- accuracy-only (gsm8k)
+# Reference gsm8k (vLLM CI gsm8k harness, 5-shot, 1319 questions): 0.834
+# --limit-mm-per-prompt {"image":0} disables the vision path (gsm8k is text-only)
+# and avoids the multimodal image-fetch init on this Mistral vision model.
+name: devstral_small_2_24b-b200
+gpu: B200
+num_gpus: 1
+nightly: true
+
+vllm:
+  model: mistralai/Devstral-Small-2-24B-Instruct-2512
+  serve_args: >-
+    --tensor-parallel-size 1
+    --enforce-eager
+    --max-model-len 4096
+    --limit-mm-per-prompt {"image":0}
+
+lm_eval:
+  model_args:
+    tokenized_requests: false
+    tokenizer_backend: null
+    timeout: 6000
+  tasks:
+    - name: gsm8k
+      num_fewshot: 5
+      model_args:
+        num_concurrent: 64
+        max_length: 4096
+        max_gen_toks: 512
diff --git a/workloads/gemma_4_31b_nvfp4_b200.yaml b/workloads/gemma_4_31b_nvfp4_b200.yaml
new file mode 100644
index 0000000..488bbfc
--- /dev/null
+++ b/workloads/gemma_4_31b_nvfp4_b200.yaml
@@ -0,0 +1,26 @@
+# Gemma-4-31B-IT NVFP4 on B200 (1xB200, TP=1) -- accuracy-only (gsm8k)
+# Reference gsm8k (vLLM CI gsm8k harness, 5-shot, 1319 questions): 0.702
+name: gemma_4_31b_nvfp4-b200
+gpu: B200
+num_gpus: 1
+nightly: true
+
+vllm:
+  model: nvidia/Gemma-4-31B-IT-NVFP4
+  serve_args: >-
+    --tensor-parallel-size 1
+    --enforce-eager
+    --max-model-len 4096
+
+lm_eval:
+  model_args:
+    tokenized_requests: false
+    tokenizer_backend: null
+    timeout: 6000
+  tasks:
+    - name: gsm8k
+      num_fewshot: 5
+      model_args:
+        num_concurrent: 64
+        max_length: 4096
+        max_gen_toks: 512
diff --git a/workloads/llama_3_3_70b_nvfp4_b200.yaml b/workloads/llama_3_3_70b_nvfp4_b200.yaml
new file mode 100644
index 0000000..b05b232
--- /dev/null
+++ b/workloads/llama_3_3_70b_nvfp4_b200.yaml
@@ -0,0 +1,26 @@
+# Llama-3.3-70B-Instruct NVFP4 on B200 (1xB200, TP=1) -- accuracy-only (gsm8k)
+# Reference gsm8k (vLLM CI gsm8k harness, 5-shot, 1319 questions): 0.911
+name: llama_3_3_70b_nvfp4-b200
+gpu: B200
+num_gpus: 1
+nightly: true
+
+vllm:
+  model: nvidia/Llama-3.3-70B-Instruct-NVFP4
+  serve_args: >-
+    --tensor-parallel-size 1
+    --enforce-eager
+    --max-model-len 4096
+
+lm_eval:
+  model_args:
+    tokenized_requests: false
+    tokenizer_backend: null
+    timeout: 6000
+  tasks:
+    - name: gsm8k
+      num_fewshot: 5
+      model_args:
+        num_concurrent: 64
+        max_length: 4096
+        max_gen_toks: 512
diff --git a/workloads/mistral_small_4_119b_nvfp4_b200.yaml b/workloads/mistral_small_4_119b_nvfp4_b200.yaml
new file mode 100644
index 0000000..fef0c90
--- /dev/null
+++ b/workloads/mistral_small_4_119b_nvfp4_b200.yaml
@@ -0,0 +1,29 @@
+# Mistral-Small-4-119B-2603 NVFP4 on B200 (1xB200, TP=1) -- accuracy-only (gsm8k)
+# Reference gsm8k (vLLM CI gsm8k harness, 5-shot, 1319 questions): 0.905
+# --limit-mm-per-prompt {"image":0} disables the vision path (gsm8k is text-only)
+# and avoids the multimodal image-fetch init on this Mistral vision model.
+name: mistral_small_4_119b_nvfp4-b200
+gpu: B200
+num_gpus: 1
+nightly: true
+
+vllm:
+  model: mistralai/Mistral-Small-4-119B-2603-NVFP4
+  serve_args: >-
+    --tensor-parallel-size 1
+    --enforce-eager
+    --max-model-len 4096
+    --limit-mm-per-prompt {"image":0}
+
+lm_eval:
+  model_args:
+    tokenized_requests: false
+    tokenizer_backend: null
+    timeout: 6000
+  tasks:
+    - name: gsm8k
+      num_fewshot: 5
+      model_args:
+        num_concurrent: 64
+        max_length: 4096
+        max_gen_toks: 512