SearchBook a call

BULLETIN / Mistral AI

Mistral Medium 3.5: review

Mistral's open-weight flagship is a practical European self-hosting option, but independent tests put it far behind the frontier leaders.

Choose it for control and European hosting, not for top-end quality.

Maker
Mistral AI
Released
28 Apr 2026
Licence
Modified MIT
Access
Mistral API, Le Chat, Mistral Studio, open weights
Context
256k tokens
Price
$1.5 input / $7.5 output per million tokens (USD)

Good at

  • Open weights you can host yourself
  • European vendor and hosting
  • Function calling and structured output
  • A wide family of Apache-2.0 smaller models

Bad at

  • Low independent index scores
  • Priced high for its measured quality
  • Licence exception for large companies
  • Frequent model retirements

Use it for

  • Organisations that must self-host a capable model
  • European teams with data residency needs
  • Builders who want small Apache-2.0 models from the same vendor

Avoid it for

  • Hard reasoning and maths
  • Long autonomous coding runs
  • Buyers optimising only for quality per dollar

Only for specific needs. Mistral Medium 3.5 is an open-weight model you can run on a handful of graphics processing units (GPUs), from a European vendor, and that is its real value. On independent quality measures it scores far below the leading closed models and below several Chinese open models, so pick it for control and hosting, not for raw capability.

First principles

Every model choice trades three things: quality, price, and control over where the model runs and who sees the data. By the independent numbers below, Mistral Medium 3.5 is weak on the first two and strong on the third. The rest of this review follows from that trade.

Memory follows parameter count. In a dense model every parameter is loaded and used for every token. The weights take roughly the parameter count times the bytes per parameter, so at 16 bits, 128B parameters is about 256GB. Mistral says four GPUs are enough; if those are 80GB cards, they hold 320GB, which leaves little room for the memory a long context needs on top of the weights. That is why we suggest sizing hardware from your own test, not from the vendor’s minimum.

Owning the weights moves the risk to you. With the weights on your own servers, the vendor cannot retire your model, change its price or see your data. In exchange you run the hardware, the security and the upgrades. Mistral’s retirement schedule, covered below, shows what users of its hosted application programming interface (API) face instead.

Price only means something next to quality. A per-token price alone says little. The useful measure is what you pay for each unit of measured capability, and on that measure several rivals do better, as the sections below show.

What it is

Mistral Medium 3.5 is Mistral AI’s flagship. Its model card gives a release date of 28 April 2026; Mistral’s announcement post is dated 22 May 2026. It is a dense model of about 128B parameters that reads text and images and returns text. The context window is 256k tokens. Mistral publishes the weights under a Modified Massachusetts Institute of Technology (MIT) licence and says the model can be self-hosted on as few as four GPUs.

It is the default in Mistral’s Le Chat assistant and has replaced Devstral 2 in the Vibe coding command-line tool.

The family around it:

  • Mistral Small 4 (March 2026, Apache 2.0) has 119B total and 6B active parameters in a mixture-of-experts design. It merges Mistral’s former reasoning, coding and vision models. It costs $0.15 input and $0.60 output per million tokens.
  • Mistral Large 3 (December 2025, Apache 2.0) has 675B total and 41B active parameters and costs $0.5 input and $1.5 output.
  • Ministral 3 comes in 3B, 8B and 14B sizes, all Apache 2.0 and all with vision. The 8B costs $0.15 per million tokens for both input and output.

Where it is good

Control. The weights are downloadable. Mistral says the model runs on four GPUs, although the serving example on its model card splits it across eight, so plan hardware from your own test. For regulated organisations that cannot send data to an American or Chinese API, that matters more than a leaderboard place.

Tool use. The model card lists structured outputs, function calling, document question answering, batching, built-in tools and predicted outputs.

Coding. Mistral reports 77.6% on the software engineering benchmark SWE-Bench Verified. This is vendor-reported, and we found no current independent SWE-bench result to confirm it.

Speed. Artificial Analysis measured about 166 output tokens per second.

The family. Few vendors offer Apache 2.0 models at 3B, 8B, 14B, 119B and 675B. For teams that want one vendor across edge devices and servers, that range is useful.

Where it is bad

Independent scores are low. Artificial Analysis gave Mistral Medium 3.5 an Intelligence Index of 14. For comparison, it gave Gemma 4 31B 19, gpt-oss-120b 12 and Claude Opus 5.5 58. Mistral Small 4 scored 11.

Price for quality. At $1.5 input and $7.5 output, it is priced high for an Intelligence Index of 14. DeepSeek-V4.1-Flash (39) and Gemini 3.8 Flash (41) scored much higher on the same index.

Licence. The Modified MIT licence includes an exception for companies with large revenue. We could not confirm the exact threshold, so check it before building a business on the weights.

Churn. Mistral retired Magistral Medium 1.2, Magistral Small 1.2 and Devstral 2 on 31 July 2026, and Mistral Medium 3.1 and 3 on 31 August 2026. Plan for model changes.

Price and access

Model Input per million Output per million Licence
Mistral Medium 3.5 $1.5 $7.5 Modified MIT
Mistral Large 3 $0.5 $1.5 Apache 2.0
Mistral Small 4 $0.15 $0.60 Apache 2.0
Ministral 3 8B $0.15 $0.15 Apache 2.0

Who should use it

On this evidence Mistral Medium 3.5 fits organisations whose first requirement is running a capable model on their own servers, or keeping a European supplier. It is a poor fit where quality per dollar is the only test. Note that Mistral Large 3 is cheaper per token on the API, and for small local jobs Ministral 3 is worth testing alongside Gemma 4 and IBM Granite 4.2.

Sources