Mistral Medium 3.5

Mistral's middle child finally comes with weights. Mistral Medium 3.5, released April 29, 2026, is a 128-billion-parameter dense model you can download from Hugging Face today: no gate, no access form, no research-only clause. It posts 77.6% on SWE-bench Verified, takes text and images, carries a 256K context window, and ships under a modified MIT license that is free for everyone below a $20 million monthly revenue bar. If you skipped every "Mistral Medium" headline because the last one was API-only, this is the release worth reading about.

What changed: Mistral's mid model is now open weights

The original Mistral Medium 3 (May 2025) was a closed, API-only model. No parameter count, no weights, and it is now deprecated, with retirement scheduled for August 31, 2026. Medium 3.5 replaces it and flips the distribution model: the full weights sit in the official mistralai Hugging Face repo, in both FP8 (about 134 GB across three safetensors shards) and BF16 (about 250 GB). We opened the file tree and checked: tokenizer, config, LICENSE, all there, nothing gated.

That matters because Medium 3.5 is not a consolation-prize model. It is positioned for agentic and coding work, with a configurable per-request reasoning effort, and its official numbers land within a few points of the best open coding models on the planet.

The Mistral naming decoder

Mistral's lineup names are a trap, and we fell into it ourselves while researching this piece. Here is the map so you don't have to reconstruct it:

Yes, that means "Medium 3" and "Medium 3.5" sit on opposite sides of the open-weights line. If you see a blog calling Medium 3 an open model, close the tab.

What it actually is

Medium 3.5 (version code 26.04) is a dense transformer: all 128B parameters are active on every token, unlike the Mixture-of-Experts design in Large 3. It is multimodal, taking text and images as input, with a vision encoder Mistral trained from scratch to handle variable image sizes and aspect ratios. The context window is 256K tokens, and you can dial reasoning effort up or down per request instead of choosing between separate instruct and reasoning variants.

Community reporting says Medium 3.5 merges what used to be separate specialist models (the Devstral coding line and Magistral reasoning line) into one set of weights. Mistral's own card does not state that, so treat it as plausible background, not fact.

The license: MIT with a $20 million asterisk

The LICENSE file in the repo is standard MIT plus one carve-out, quoted directly: "You are not authorized to exercise any rights under this license if the global consolidated monthly revenue of your company (or that of your employer) exceeds $20 million." Above that bar, you email Mistral for a commercial license.

Practical reading: $20 million per month is roughly $240 million a year in revenue. Every indie hacker, startup, research lab, and mid-size company on earth is below it. You can use it commercially, fine-tune it, and redistribute it. It is still worth one critical note: within Mistral's own family, the smallest model (Small, Apache 2.0) and the biggest (Large 3, Apache 2.0) are more permissively licensed than this middle one. Mistral put the revenue cap exactly where the enterprise money is.

Benchmarks: official numbers first

Mistral's model card states two headline results in plain text: 77.6% on SWE-bench Verified and 91.4% on the tau3-Telecom agentic benchmark. The rest of the card's comparisons ship as chart images, so exact MMLU and GPQA digits are not published as numbers anywhere we could verify.

How does 77.6% stack up? Here is the open-weights coding field as reported by community trackers and roundups (these cross-model numbers are community-reported, not Mistral's):

The independent Artificial Analysis Intelligence Index scores Medium 3.5 at 30, second of 62 models tracked at the time of writing. Third-party API benchmarks measured about 102 tokens per second on La Plateforme, where pricing is $1.50 per million input tokens and $7.50 per million output.

The honest summary: the giant MoE models still hold the coding crown by about three points, but Medium 3.5 gets you within reach of them from a single-node, dense, self-hostable package.

Run it locally, if your hardware can

Let's be direct: 128B dense is not a laptop model. Dense means every token streams all 128B weights, so your tokens per second are roughly memory bandwidth divided by model size. That is why MoE models fly on the same box while this one crawls. Mistral says it self-hosts on as few as 4 GPUs, which in practice means 80GB-class cards for the FP8 weights. The one measured community datapoint we found: two RTX PRO 6000 Blackwell cards (192GB total) running FP8 under vLLM deliver 26 to 35 tokens per second on prose and 37 to 43 on code. Below that tier, the community GGUF quants from unsloth are the realistic path:

There is an official Ollama library entry, so the setup is one line if your machine qualifies:

# Ollama (needs ~80GB free RAM/VRAM for the default quant)
ollama run mistral-medium-3.5:128b

# Self-host FP8 with vLLM (official recipe; needs a vLLM NIGHTLY build as of July 2026)
vllm serve mistralai/Mistral-Medium-3.5-128B --tensor-parallel-size 8 \
  --tokenizer_mode mistral --config_format mistral --load_format mistral \
  --enable-auto-tool-choice --tool-call-parser mistral --reasoning-parser mistral

# No big iron? The hosted API. Set MISTRAL_API_KEY first.
curl https://api.mistral.ai/v1/chat/completions \
  -H "Authorization: Bearer $MISTRAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"mistral-medium-3-5","messages":[{"role":"user","content":"Refactor this function and explain why."}]}'

Limitations and gotchas

• Dense means no free lunch. A 128B MoE with 40B active would run far faster on the same hardware. Every token here touches all 128B weights.
• The license is not Apache 2.0. The $20M/month carve-out is irrelevant for most of us but makes lawyers at big companies read the file twice.
• Mistral's chart-image benchmarks mean key numbers (MMLU, GPQA) are unverifiable in text form. We flagged what is official versus community-reported above.
• It trails DeepSeek V4 and Kimi K2.6 on SWE-bench by about three points, per community trackers. If leaderboard-max coding is the only goal, those remain ahead (and much heavier).
• Early GGUFs shipped with a broken YaRN long-context config (repetition and forgetting on long chats). It is fixed upstream, so if you downloaded a quant in the first days after launch, re-download it.
• Vision does not work through Ollama (the GGUF needs a separate mmproj file); use llama.cpp for image input, with a BF16 or F32 mmproj, not F16.
• vLLM support is nightly-only at the time of writing, and fine-tuning a 128B dense model is out of QLoRA range for typical rigs. If you want to fine-tune a Mistral, use Small 3.1.

Who should use it

Pick Medium 3.5 if you want near-frontier agentic coding from weights you control, and you have either a multi-GPU node, a 128GB unified-memory machine, or a tolerance for API pricing. It is the strongest open Mistral for coding and agent work, sitting between the local-friendly Small 3.1 and the chat-focused Large 3. If your box tops out at 24GB of VRAM, run Small 3.1 locally and call Medium 3.5 over the API when the task deserves it.

Sources and further reading

• Mistral: Medium 3.5 announcement
• Official weights repo (Hugging Face)
• Mistral docs: Medium 3.5 model card
• unsloth GGUF quants
• Ollama library entry
• Artificial Analysis: independent benchmarks

Tested on: not independently benchmarked. A 128B dense model exceeds our local bench, so all numbers above come from Mistral's model card (marked official) or community trackers (marked community-reported), with sources linked.
Date checked: 2026-07-10