Skip to content
Bud's Wiki

Mistral Medium 3.5 128B

TL;DR

Section titled “TL;DR”

Mistral Medium 3.5 128B is Mistral’s first flagship “merged” model — a dense 128B-parameter checkpoint with a 256k context window that replaces both Mistral Medium 3.1 (instruct) and Magistral (reasoning) in Le Chat, and Devstral 2 (coding) in the Vibe CLI. A per-request reasoning_effort parameter toggles between fast instant-reply mode and step-by-step reasoning. The vision encoder was trained from scratch to handle variable image sizes and aspect ratios. Released under a Modified MIT License (open-source except for companies with large revenue), with day-0 vLLM / SGLang / llama.cpp / Ollama / Transformers support and a separately released EAGLE speculative-decoding head.

Key claims

Section titled “Key claims”
  • The model unifies instruct, reasoning, and coding capabilities into a single dense 128B checkpoint, replacing Mistral Medium 3.1 + Magistral in Le Chat and Devstral 2 in the Vibe coding agent [§Mistral Medium 3.5 128B].
  • Dense (not MoE) architecture at 128B parameters with 256k context length, multimodal text+image input with text output [§Key Features].
  • Per-request reasoning_effort parameter toggles between 'none' (fast reply) and 'high' (reasoning mode with test-time compute); recommended sampling is temp=0.7 / top_p=0.95 for 'high' and temp ≤ 0.7 / top_p=None for 'none' [§Recommended Settings].
  • The vision encoder was trained from scratch to handle variable image sizes and aspect ratios [§Mistral Medium 3.5 128B].
  • Agentic benchmarks: 91.4% on τ³-Telecom and 77.6% on SWE-Bench Verified, claimed to supersede Devstral across all benchmarks [§Agentic Benchmarks].
  • A separately released EAGLE speculative-decoding head is provided to accelerate local inference under vLLM and SGLang [§vLLM Recommended, §SGLang].
  • Released under a Modified MIT License: open-source for commercial and non-commercial use, with exceptions for companies with large revenue [§License].
  • Day-0 multi-backend support shipped at release: vLLM (recommended), SGLang (with dedicated H100/H200 and B200/B300 docker tags), llama.cpp via Unsloth GGUFs, Ollama, Transformers via Mistral3ForConditionalGeneration, and the Mistral Vibe CLI [§Inference].
  • A bug warning: the original Transformers config had an incorrect entry that caused long-context performance degradation; this has been fixed in a specific commit, and GGUFs generated from the pre-fix config are also affected [§Mistral Medium 3.5 128B].

Method

Section titled “Method”

The model card is product-facing and does not include a technical report or full architecture description. The visible architectural picks are: a dense 128B decoder-only transformer (no MoE), 256k context, multimodal vision+text input via a from-scratch vision encoder with native variable-resolution / variable-aspect-ratio support, and a per-request reasoning toggle implemented in the chat template (the reasoning_effort argument passed to apply_chat_template or to the OpenAI-compatible chat completion API). The reasoning trace is emitted between [THINK] [/THINK] tags when reasoning is on, mirroring the same convention used by Mistral Small 4 (Mistral Small 4 119B (instruct + reasoning + Devstral unified MoE)) and Magistral. Tool calling is native, with --tool-call-parser mistral and --reasoning-parser mistral as the recommended vLLM flags; the released vLLM recipe specifies --tensor-parallel-size 8, suggesting the model is sized to fit on an 8×H100/H200 node at deployment. Companion artifacts: an EAGLE speculative-decoding head for accelerated local inference, Unsloth GGUF quantizations, and the Mistral Vibe CLI as the recommended agentic-coding harness.

Results

Section titled “Results”

The card reports two headline agentic-benchmark numbers — 91.4% on τ³-Telecom and 77.6% on SWE-Bench Verified — and claims Medium 3.5 supersedes Devstral across all benchmarks the team measured, which is what motivates replacing Devstral 2 with Medium 3.5 in the Vibe CLI. The “Instruction Following, Reasoning, and Coding Benchmarks” section is described as comparing Medium 3.5 with competing models on the unified surface (IFEval-style + math + code), but the specific tables / numbers are present only as embedded figures in the HF model card and are not transcribed in the body. The HF evaluation results panel surfaces a SWE-bench Verified number consistent with the 77.6% figure. No throughput, latency, or VRAM-footprint numbers are provided.

Why it’s interesting

Section titled “Why it’s interesting”

This is the second Mistral flagship release in the wiki to fold instruct + reasoning + code into a single checkpoint with a per-request reasoning toggle — following Mistral Small 4 119B (instruct + reasoning + Devstral unified MoE), which did the same thing at 119B-total / 6.5B-activated MoE scale. The architectural choice is the interesting contrast: Medium 3.5 is dense 128B, whereas Small 4 is MoE; both ship the same reasoning_effort UX and [THINK] [/THINK] trace convention, so Mistral now has a coherent dense-vs-MoE pair at adjacent parameter scales using identical inference contracts. The packaging continues to fit the deep-release-package pattern catalogued in Open foundation-model releases: base checkpoint + EAGLE speculative head + day-0 vLLM/SGLang/llama.cpp/Ollama/Transformers serving + a custom CLI (Mistral Vibe) as the reference agentic harness. The “merged model” framing — explicitly replacing three separately-released model families with one weight set — is becoming the modal release shape in 2026, also seen in GPT-5.3 Instant (GPT-5.3 Instant: Smoother, more useful everyday conversations) and Step 3.5 Flash (Step 3.5 Flash: Open Frontier-Level Intelligence with 11B Active Parameters).

See also

Section titled “See also”