Mistral Small 4 replaces three separate Mistral models - Magistral for reasoning,
Devstral for coding agents, and Mistral Small for instruct - with a single 119B MoE
model (128 experts, 4 active, 6.5B active params per token). You pick the behavior
per request with a reasoning_effort parameter:
reasoning_effort="none": fast chat-style responses, equivalent to Mistral Small 3.2reasoning_effort="high": deep step-by-step reasoning, equivalent to Magistral
Same weights, same deployment, different behavior at inference time.
The efficiency numbers are notable: it matches or beats GPT-OSS 120B on LCR, LiveCodeBench, and AIME 2025 while producing 20-75% shorter outputs than comparable Qwen reasoning models at similar scores. Other specs: 256k context window, multimodal (text + image input), Apache 2.0.
NVFP4 quantized checkpoint
Mistral also released an NVFP4 (4-bit float) quantized version
(Mistral-Small-4-119B-2603-NVFP4), created with llm-compressor in collaboration
with vLLM and Red Hat, with NVIDIA contributing kernel optimizations for vLLM and
SGLang. It uses post-training-activation quantization, so accuracy stays close to the
BF16 base while significantly reducing memory footprint and improving throughput.
The NVFP4 checkpoint uses a different attention backend flag when serving:
# Base model
vllm serve mistralai/Mistral-Small-4-119B-2603 \
--attention-backend FLASH_ATTN_MLA \
--tensor-parallel-size 2 ...
# NVFP4 quantized
vllm serve mistralai/Mistral-Small-4-119B-2603-NVFP4 \
--attention-backend TRITON_MLA \
--tensor-parallel-size 2 ...
Both models require a custom vLLM build until the Mistral parsing fixes land in main
(expected mid-March 2026). Mistral provides a Docker image (mistralllm/vllm-ms4:latest)
with everything pre-patched.