What it is
Mistral Medium 3.5 is a dense 128-billion-parameter model announced on May 22, 2026 and still labelled a public preview. It is Mistral's first flagship "merged" model: one set of weights covers chat, reasoning and coding, replacing Mistral Medium 3.1 and Magistral in Le Chat and Devstral 2 in the Vibe coding agent. It reads text and images, writes text, and has a 256K-token context, with reasoning switched on or off per request.
The weights are on Hugging Face under a modified MIT licence that gives no rights to any company whose monthly revenue exceeds $20 million; those companies need a commercial licence from Mistral or its hosted services. Mistral says the model can be self-hosted on as few as four GPUs. The API price is $1.50 per million input tokens and $7.50 per million output tokens.
It is the default model in Le Chat and Mistral Vibe, powers Vibe's remote coding agents, and is available through Mistral's API, OpenRouter and as an NVIDIA NIM container.
Our review
Mistral Medium 3.5 is a strong coding model sized for one server, and its licence decides who may run it. The weights ship under a modified MIT licence that withdraws every right from a company whose consolidated monthly revenue topped $20 million in the previous month, and the restriction follows derivatives and third-party hosts. Above that line the choices are a commercial licence granted at Mistral's discretion, or Mistral's own API.
Mistral reports 77.6% on SWE-Bench Verified, ahead of the Devstral 2 it replaces and of Qwen3.5 397B A17B, and 91.4 on τ³-Telecom. Those are vendor figures against a spring 2026 field, and the rest are published only as charts. The dense design is the practical point: most weights ship in FP8, about 134 GB in all, which is how Mistral can claim four GPUs. Its own vLLM example on the model card uses eight.
On the API the arithmetic has turned against it. At $1.50 input and $7.50 output per million tokens it costs more than Mistral Large 4, a 1-trillion-parameter mixture-of-experts model previewed on 6 October at $1.36 and $4.18 with a 1M-token context against Medium's 256K. Both are still previews. Medium's remaining edge is that its weights are downloadable now and fit a single node; Large 4's are promised for the end of October and will need far more hardware.
Teams already self-hosting should check their files before judging quality. Mistral warns that the original Transformers config carried an incorrect entry that degraded long-context performance, and that GGUFs converted before the fix inherit it. A community quantisation pulled in the first weeks after launch may be measurably worse than the model Mistral benchmarked.
Where it earns its place
A 60-person European fintech with a platform team and its own GPU servers
Run a coding agent over private repositories without code leaving its infrastructure
Serve the weights with vLLM on four 80 GB GPUs and point Mistral Vibe at the local endpoint, a setup the model card documents. Below the $20 million monthly revenue line the licence costs nothing; the spend is hardware and the engineer who keeps it running.
An engineering manager whose team already pays for Le Chat Team or Enterprise
Hand refactors, test generation and dependency upgrades to background agents
Medium 3.5 powers Vibe remote agents, which run each session in an isolated sandbox and open a GitHub pull request when done. On Vibe, inputs train Mistral's models unless someone opts out; only Enterprise accounts are opted out by default.
A procurement lead at a retailer with more than $20 million in monthly revenue
Decide whether the open weights can be used at all
They cannot without a separate commercial licence, and the bar covers fine-tunes and third-party hosts too. The practical routes are Mistral's API at $1.50/$7.50 per million tokens, or Large 4, which is cheaper per token and has four times the context.
Before you commit
- Setup
- API model ID mistral-medium-latest. Self-hosting needs vLLM nightly or SGLang images for H100/H200 and B200/B300, plus the corrected config; an EAGLE draft model speeds up decoding.
- Integrations
- Chat completions with tool calls, JSON output and reasoning_effort set to none or high; available on Mistral's API, OpenRouter, NVIDIA NIM and build.nvidia.com.
- Data and privacy
- Modified MIT: usable only if your company's monthly revenue is under $20 million. On Mistral Studio, free-mode inputs and outputs may be used for training, pay-as-you-go customers can opt out, and Vibe trains by default except on Enterprise.
- Cost at scale
- 1B input and 100M output tokens a month is about $2,250 at list, or about $1,125 with the 50% batch discount; the same load on Large 4 is about $1,778. Self-hosting means about 134 GB of mostly FP8 weights, so four 80 GB GPUs as a floor.
- Limits
- 256K context; still a public preview. It replaced Medium 3.1 (retired 31 August 2026) and Devstral 2 (retired 31 July 2026), so older model IDs no longer work.
What's good
- Dense 128B weights, mostly FP8, fit four 80 GB GPUs
- 77.6% on SWE-Bench Verified (vendor-reported)
- Image input and a per-request reasoning switch
- Default model for Vibe remote agents and Le Chat
What's not
- Licence bars companies above $20M monthly revenue
- API costs more per token than Mistral Large 4
- 256K context, a quarter of Large 4's
- Early GGUFs carry a long-context config bug
Best for
Companies under $20 million in monthly revenue that want to self-host a strong coding and agent model on one four-GPU server.
Alternatives
Mistral Medium 3.5: common questions
- Is Mistral Medium 3.5 worth it?
- A capable four-GPU coding model whose licence excludes large companies and whose API price now loses to Mistral's own Large 4. We rate it 3.8 out of 5.
- How much does Mistral Medium 3.5 cost?
- Open weights under a modified MIT licence (no rights for companies above $20M monthly revenue). API: $1.50 per 1M input tokens, $7.50 per 1M output; batch 50% off.
- What are the downsides of Mistral Medium 3.5?
- Licence bars companies above $20M monthly revenue; aPI costs more per token than Mistral Large 4; 256K context, a quarter of Large 4's; early GGUFs carry a long-context config bug.
- Who is Mistral Medium 3.5 best for?
- Companies under $20 million in monthly revenue that want to self-host a strong coding and agent model on one four-GPU server.
- How much does Mistral Medium 3.5 cost for a team or in production?
- 1B input and 100M output tokens a month is about $2,250 at list, or about $1,125 with the 50% batch discount; the same load on Large 4 is about $1,778. Self-hosting means about 134 GB of mostly FP8 weights, so four 80 GB GPUs as a floor.
- Is Mistral Medium 3.5 safe for company data?
- Modified MIT: usable only if your company's monthly revenue is under $20 million. On Mistral Studio, free-mode inputs and outputs may be used for training, pay-as-you-go customers can opt out, and Vibe trains by default except on Enterprise.
- What are the best Mistral Medium 3.5 alternatives?
- The closest alternatives we have reviewed are Mistral Large 4, Mistral Small 4, Qwen3.8-Max and DeepSeek-V4.1-Flash.
Answers checked Oct 22, 2026.