Mistral Small 4 119B (instruct + reasoning + Devstral unified MoE)

TL;DR

Mistral Small 4 119B 2603 is an Apache-2.0 open-weights MoE that unifies three previously-separate Mistral model families — Instruct, Reasoning (formerly Magistral), and Devstral (code) — into a single checkpoint. The architecture is 119B total / 6.5B activated parameters across 128 experts (4 active), with 256k context and native multimodal (text + image → text) input. A per-request reasoning_effort switch toggles between fast instant-reply mode and step-by-step reasoning. Mistral ships the base checkpoint alongside an Eagle speculative-decoding head, an NVFP4-quantized variant, and day-0 support across vLLM / llama.cpp / LM Studio / SGLang / Transformers.

Key claims

The model unifies Instruct + Reasoning (Magistral) + Devstral capabilities into one checkpoint with a per-request reasoning toggle, replacing the previous three separate model families [§Mistral Small 4 119B A6B].
Architecture: 128-expert MoE with 4 experts active per token, 119B total parameters, 6.5B activated per token, 256k context length, multimodal text+image input [§Key Features].
Versus Mistral Small 3, the latency-optimized setup is ~40% faster end-to-end and the throughput-optimized setup serves ~3× more requests per second [§Mistral Small 4 119B A6B].
A trained Eagle speculative-decoding head and an NVFP4 4-bit float quantized checkpoint are released as separate sibling repos at launch [§Mistral Small 4 119B A6B].
The released vLLM serving recipe specifies --attention-backend FLASH_ATTN_MLA, implying Multi-head Latent Attention (DeepSeek-style) is the attention design [§Serve the Model].
With reasoning_effort="high", Mistral Small 4 reportedly matches or beats GPT-OSS 120B on three benchmarks while emitting shorter outputs — e.g. AA LCR 0.72 at ~1.6K characters of output vs Qwen models requiring 3.5–4× more output (5.8–6.1K) for comparable scores, and ~20% less output than GPT-OSS 120B on LiveCodeBench [§Comparison with other models].
GPQA Diamond score reported on the HF evaluation panel: 71.2 [§Evaluation results].
Released under Apache 2.0 for commercial and non-commercial use [§License].

Method

The release page does not include a technical report — only the model card. From the card, the architectural picks are: a granular Mixture-of-Experts (128 experts / 4 active) at 119B total / 6.5B active, native multimodal (vision + text in, text out) via the Mistral3ForConditionalGeneration Transformers class, 256k context, and per-request reasoning toggling via a reasoning_effort argument that flips the model between a Mistral-Small-3.2-style fast reply mode and a Magistral-Small-style step-by-step reasoning mode (with reasoning trace emitted between [THINK] [/THINK] tags). Tool calling and JSON output are native, and the recommended vLLM launch flags include --tool-call-parser mistral, --reasoning-parser mistral, and FLASH_ATTN_MLA as the attention backend (which signals MLA-style KV-cache compression analogous to DeepSeek’s approach). Companion artifacts: an Eagle speculative-decoding head trained for the model, and an NVFP4 quantized variant for 4-bit float inference.

Results

The card’s “Comparison with other models” section reports that at reasoning_effort="high", Mistral Small 4 matches or surpasses GPT-OSS 120B on three benchmarks while producing significantly shorter outputs: on AA LCR it scores 0.72 with ~1.6K output characters where Qwen comparators require 5.8–6.1K (3.5–4×) for similar scores; on LiveCodeBench it produces ~20% less output than GPT-OSS 120B. The HF “Evaluation results” panel lists a GPQA Diamond score of 71.2. No numbers are reported for the non-reasoning mode in the card body. No throughput numbers are given in absolute terms, only the ratios vs Mistral Small 3 (40% latency reduction; 3× higher RPS).

Why it’s interesting

The headline architectural choice — a single MoE that subsumes the previously-separate Instruct / Reasoning / Code model families via a per-request reasoning toggle — matches the broader 2026 trend of merging “fast” and “thinking” modes into one checkpoint, also seen in GPT-5.3 Instant: Smoother, more useful everyday conversations and the Step 3.5 Flash release (Step 3.5 Flash: Open Frontier-Level Intelligence with 11B Active Parameters). The packaging — base checkpoint + Eagle speculative head + NVFP4 quantized variant + day-0 multi-backend serving — extends the “deep release package” pattern documented in Open foundation-model releases (companion-models pattern seen in Qwen3-VL-Embedding/Reranker, LTX-2’s distilled+quantized variant grid). The FLASH_ATTN_MLA backend choice is notable: it suggests Mistral has adopted DeepSeek-style Multi-head Latent Attention for KV-cache compression, complementing the broader efficient-attention discussion in How Attention Got So Efficient [GQA / MLA / DSA].