Verdict Is It Worth It

Mistral Medium 3.5 Review: The Verdict

Europe's open-weight play ships a 256k-context, multimodal coder that self-hosts on four GPUs. Here's where it actually earns a slot in the rotation — and where it doesn't.

Abstract blue network of interconnected nodes and lines
Photo via Unsplash
The receipts
  • Mistral Medium 3.5 shipped April 28, 2026: a 128B dense multimodal model, 256k-token context, adjustable reasoning via a reasoning_effort knob.
  • Open weights under a Modified MIT license on Hugging Face — self-hostable on as few as four GPUs.
  • Posted 77.6% on SWE-Bench Verified and 91.4 on the tau-3 Telecom agentic tool-use benchmark. API runs $1.50/M in, $7.50/M out.
  • Not the cheapest and not the smartest — but the open weights plus that price is the reason it stays in my rotation.
Short answer

Mistral Medium 3.5 is a 128B dense multimodal model from Mistral AI, released April 28, 2026, with a 256k-token context window and open weights under a Modified MIT license. It scored 77.6% on SWE-Bench Verified, self-hosts on as few as four GPUs, and is priced at $1.50 per million input tokens and $7.50 per million output.

Short version of this Mistral Medium 3.5 review: it’s the right model for a specific job, not the best model in the room. Mistral shipped Medium 3.5 on April 28, 2026 as a frontier-class multimodal model tuned for agentic and coding work, with open weights you can actually download and run. That last part is the whole story.

I’ve been a model-pluralist for a while now. Right model for the job. And “European open-weight coder you can self-host” is a job that keeps coming up.

What is Mistral Medium 3.5?

Mistral Medium 3.5 is a 128B dense multimodal model from the French lab Mistral AI, released April 28, 2026, built for agentic and coding use with adjustable reasoning through a reasoning_effort parameter. Context window is 256k tokens. The weights are open under a Modified MIT license.

The vision side isn’t bolted on. Mistral trained a vision encoder from scratch to handle variable image sizes and aspect ratios, and merged it into a single 128B dense model. So it reads a screenshot of a broken UI, not just your description of it.

How good is Mistral Medium 3.5 at coding and agents?

On the numbers Mistral published, good — not chart-topping, but genuinely useful. It posted 77.6% on SWE-Bench Verified, which Mistral says beats Devstral 2 and Qwen3.5 397B A17B, and 91.4 on the tau-3 Telecom agentic tool-use benchmark. There is no official HumanEval score.

Peep that carefully. SWE-Bench Verified and tau-3 are the receipts that matter for agents — real repo fixes and real tool-calling, not toy function completion. If you see a “92% HumanEval” number floating around for this model, it’s not from Mistral. Cut it from your notes. This is exactly the kind of model I keep in the rotation next to the best open-weight coding models — a workhorse, not a hero.

How much does Mistral Medium 3.5 cost to run?

The API runs $1.50 per million input tokens and $7.50 per million output, per Mistral’s model card. That’s the meter. The bigger lever is that you don’t have to use it: open weights mean self-hosting on as few as four GPUs, with vLLM support for serving.

Four GPUs. Not an eight-GPU node, not a rack. That changes the math for anyone running heavy agent loops, because the per-token bill disappears the moment you’re on your own hardware. Same reasoning I used when I looked at GLM running local — the frontier tax is optional if the open weights are good enough for the step.

Mistral Medium 3.5 vs GPT-5.6: where does each win?

Straight up: GPT-5.6 wins the trophy, Mistral wins the deployment. OpenAI’s family shipped to general availability July 9, 2026, and the balanced Terra tier carries a ~1.05M-token context with reasoning you can’t touch on open weights. But you rent Terra. You own Medium 3.5.

ModelContextPricing (in / out per M)Open weightsStandout benchmarkWhere it loses
Mistral Medium 3.5256k$1.50 / $7.50Yes, Modified MIT77.6% SWE-Bench VerifiedNot the top reasoning score; smaller context than Terra
GPT-5.6 Terra~1.05MRented (OpenAI API)NoTop-tier reasoning, huge contextClosed weights — no self-host, no data-control story; text + image only, no audio

That “no audio” line matters if someone told you Terra is fully multimodal. Per the reference specs, Terra takes text and image in and puts text out — no native audio. Both of these models are eyes-and-text, not ears.

Where I land

Mistral Medium 3.5 is the truth for one specific play: agentic coding you want to own, self-hosted, without frontier-lab pricing dictating your architecture. It’s not the smartest model shipping in July 2026 and it’s not the cheapest. It’s the one you can put on your own four GPUs and stop asking permission.

Right model for the job. This one’s got a job. Keep it in the rotation as an open-weight anchor, run a frontier model for the step that actually needs frontier reasoning, and let the meter work for you instead of against you.

The trophy goes to the lab with the biggest cluster. The deployment goes to whoever ships. Ship.

#TheAIMogul

Bottom lineMistral Medium 3.5 is the right model for a specific job: agentic coding you want to own and self-host without frontier-lab pricing. If you need the absolute top of the reasoning charts, GPT-5.6 still wins the trophy — but you can't run that on your own four GPUs.

Frequently asked

Is Mistral Medium 3.5 open source?
The weights are open under a Modified MIT license, published on Hugging Face — not Apache 2.0, which some early write-ups got wrong. That means you can download it and self-host on as few as four GPUs, with the license carrying Mistral's own modifications. It is 'open weight,' which is the practical thing builders care about: you can run it yourself.
How much does the Mistral Medium 3.5 API cost?
Mistral's model card lists $1.50 per million input tokens and $7.50 per million output tokens. That undercuts frontier flagships on the input side while sitting well above the cheapest open-weight options. The real cost lever is self-hosting: with open weights on four GPUs, heavy agentic workloads can skip the per-token meter entirely.
How good is Mistral Medium 3.5 at coding?
The official coding number is 77.6% on SWE-Bench Verified, which Mistral says surpasses Devstral 2 and Qwen3.5 397B A17B. It also posted 91.4 on the tau-3 Telecom agentic tool-use benchmark. There is no official HumanEval score, so ignore any HumanEval figure you see quoted — it isn't from Mistral.
What's the context window on Mistral Medium 3.5?
256,000 tokens. That's enough to hold a mid-sized codebase, long agent trajectories, or a stack of documents in a single call. It's multimodal too — a vision encoder trained from scratch handles variable image sizes and aspect ratios — so it reads screenshots and diagrams, not just text.
Mistral Medium 3.5 vs GPT-5.6 — which should I use?
Different jobs. GPT-5.6's Terra tier carries a ~1.05M-token context and top-tier reasoning, but you rent it from OpenAI. Mistral Medium 3.5 gives you open weights you can self-host and own. If data control and cost predictability matter more than the last few points of benchmark, Mistral takes it.