Mistral Models: Large 4, Medium 3.5, Small 4 and Licences

Mistral AI's current flagship is Large 4: about 1 trillion parameters, 1M-token context, text and image input, open weights, and a hosted price of $0.70 input and $2.10 output per million tokens. Medium 3.5 (128B dense) is deprecated on Mistral's API from 31 October 2026 in favour of Large 4. Small 4 (119B, 6.5B active) is the Apache-2.0 option. Checked 31 October 2026.

The lineup

Everything in this table comes from Hugging Face cards, API metadata and Mistral's docs, read on 31 October 2026.

Medium 3.5's Hugging Face card calls it Mistral's first flagship merged model, handling instruction following, reasoning and coding in one set of weights. Large 4 does the same job at a larger scale, which is why Mistral's docs name it as the replacement.

Mistral Large 4

Mistral says Large 4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centres. The card claims it unifies instruction following, reasoning, coding and agentic work, with particular strength in cybersecurity, finance and law, and "state-of-the-art visual grounding among open weights". It is fluent in 160 or more languages, including every official EU language, according to the card.

Practical details worth knowing:

• Three reasoning levels. Use high for complex and agentic work, medium for everyday use, low for quick chat replies.
• Sampling. Mistral recommends temperature 1.0 and top_p 0.95 at every effort level.
• Tool use. The card advises a clear system prompt and keeping the tool list as small as the job allows.
• Speculative decoding. The checkpoint carries a DSpark draft head for faster serving.
• Quantised build. An NVFP4 checkpoint is published for Blackwell GPUs, with a smaller footprint and, Mistral says, the same performance. That is the vendor's claim, not something we tested.

Price

Mistral's model page lists Large 4 at $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens, with a 1M context. Mistral's launch page repeats the $0.70 and $2.10 figures. We checked both on 31 October 2026.

Large 4's price is close to DeepSeek V4 Pro's off-peak rate, and Mistral does not apply a peak surcharge on its page. Mistral's pricing page lists Small 4 at $0.15 input and $0.60 output per million tokens, with cached input at $0.015, and does not list Medium 3.5.

Mistral Medium 3.5

Medium 3.5 matters mostly for two reasons. It is the model Mistral says replaces Devstral 2 in its Vibe CLI coding agent, and its card reports coding numbers we can quote.

The card says it "supersedes all our previous coding models, namely Devstral, across all benchmarks". That is Mistral's statement about its own models. The SWE-bench chart on the card is an image, so we do not list comparison rows for other labs' models.

The weights remain on Hugging Face under a modified MIT licence with, per the card, exceptions for companies with large revenue. The card gives no revenue figure, and we did not read the licence text. One correction in the card is worth passing on. An earlier Transformers config entry caused long-context performance degradation and was fixed in a later commit, and GGUF files made from the old config may be affected. Re-download if you converted early.

Mistral's own docs now say the hosted model is deprecated, with 31 October 2026 as the deprecation date and Large 4 as the replacement. That is the date of this check. If you call mistral-medium-3-5 in production, plan to move to mistral-large-4, and test, since a larger model with different default effort may behave differently. The model page states the deprecation date and no separate retirement date.

Mistral Small 4

Small 4 is the model to look at if you want an Apache-2.0 Mistral. Its card describes a 119B-parameter mixture-of-experts with 128 experts, 4 active, only 6.5B parameters activated per token and a 256K context. It combines instruct, reasoning and coding in one model and targets developers working on SWE automation and codebase exploration. The card's benchmark charts are images, so no scores are quoted here.

Mistral also publishes NVFP4 and speculative-decoding (eagle) variants. Active-parameter count is small, but memory is set by the total: 119B parameters is still well beyond a single consumer GPU.

The licences compared

Large 4's licence is the clearest of the three because we could read it. It grants the usual MIT rights, then says that if you or an affiliate operate a Model-as-a-Service business, you need a separate commercial licence before any commercial use. It defines that as giving third parties inference or post-training access with meaningful control over inputs, parameters or training data. It excludes end-user products where the model sits behind specific features, and mere request relaying. Internal use, and anything through Mistral's own products, is outside the rule.

For most application builders, that means you can ship a product built on Large 4. If your business is selling model access, contact Mistral first. Compare this with the revenue-gated clauses in Kimi's and GLM's licences.

Running the weights

Plan around the download sizes. Large 4 is about 1.06 TB, or about 610 GB in the NVFP4 build, which targets Blackwell GPUs. Medium 3.5 is about 267 GB, so it needs a multi-GPU node even though it is dense. If you only have one GPU, neither is a fit; look at Qwen3.8-27B or gpt-oss-20b. For serving tools read Ollama vs vLLM.

Both big cards say to pass the mistral tool-call and reasoning parsers when serving, otherwise tool use and thinking will not parse correctly. If you serve with a custom stack, check parser support before blaming the model. If you run short of GPU memory, the Large 4 card suggests lowering max model length or the number of concurrent sequences.

Moving from Medium 3.5 to Large 4

Mistral's deprecation notice makes this move a near-term job for anyone on the hosted Medium model. The documented differences give you a checklist.

Do not assume a bigger model is a drop-in replacement. Larger MoE models can respond with longer reasoning at the default high setting, which raises output tokens. Run your evaluation set at medium and high, and compare the cost per correct answer, not just accuracy. If you self-host Medium 3.5 weights, nothing forces a change: the file is yours and the licence stays what it was.

Choosing a reasoning level

Mistral's card for Large 4 gives a simple mapping, and it is worth following before you invent your own. Use low for quick chat replies where a wrong answer is cheap. Use medium for day-to-day work, such as explaining code or drafting a change. Use high for complex prompts and agentic runs, which is also the default. The same weights serve all three, so you can route traffic by task: a support bot on low, a code review agent on medium and a long refactor on high.

Routing like that is a cost lever as well as a speed lever. Output tokens cost three times input tokens at Large 4's prices ($2.10 against $0.70), and reasoning is output. Sending easy requests to low effort cuts the expensive side of the bill. We have no measured token counts per effort level to quote.

Pointing Mistral Vibe at your own server

The Large 4 and Medium 3.5 cards show the same pattern for running Mistral's coding agent against a local vLLM server. You serve the model with the Mistral tool-call and reasoning parsers, install the Vibe tool, and add a model entry to its config file. In outline:

active_model = "mistral-large-4"
# provider entry points at your local OpenAI-compatible server
# max_context_length = 1048576 for Large 4

That is a sketch of the card's structure, not a complete file. The card's provider block is [[providers]] with name = "vllm", api_base = "http://<your-host-url>:8000/v1", api_key_env_var = "", backend = "generic" and api_style = "reasoning", and a matching [[models]] entry is also needed. Keep exactly one active_model entry, as the card warns. If the server runs on another machine, treat it like any internal API: put it behind authentication, since vLLM does not secure itself by default.

How the 1M context should be used

A million tokens is roughly the size of a mid-sized codebase, but the card stresses the practical limits. It says the checkpoint supports the 1M window and, in the same breath, suggests lowering the maximum model length if memory is short. Long context is paid for in key-value cache, and your concurrency drops as the window grows. Start by serving 128K or 256K and raise it when a real task needs more. For retrieval-style questions over many files, a search tool plus a smaller window is usually cheaper than stuffing everything into one prompt, though we have no Mistral-specific measurement to quote.

Where it is served

• Mistral's API (La Plateforme's successor, called Studio in Mistral's docs): Large 4, Medium 3.5 until retirement and Small 4. Mistral's model catalogue also lists third-party open models, including Z.ai GLM 5.3, GLM 5.3 Flash and DeepSeek V4.1 Flash 1017, which makes it a single place to try them.
• Cloud marketplaces and hosts. AWS Bedrock, Hugging Face Inference and others may carry Mistral models, but we did not confirm Large 4 on any.
• Your own GPUs via Runpod, Modal or Lambda.

Coding agents

Mistral's own agent is Vibe, available as a CLI, IDE integration and on the web. The Large 4 and Medium 3.5 cards both show how to point Vibe at a local vLLM server: install the mistral-vibe tool, set the active model in ~/.vibe/config.toml and give it the 1,048,576 context length for Large 4. Mistral says Medium 3.5 replaced Devstral 2 as Vibe's model. For third-party agents with a custom endpoint see OpenCode, Cline and Claude Code. Whether each handles Mistral's tool-call format cleanly is not documented on the cards. For self-hosting, the Large 4 card gives vLLM flags --tool-call-parser mistral, --enable-auto-tool-choice and --reasoning-parser mistral, and calls the tool-call parser required for tool usage.

Who should pick Mistral

Limits and gotchas

• Deprecation. Medium 3.5 is deprecated on the hosted API as of this check.
• Reasoning default. Large 4 defaults to high effort, which is slower and more expensive in tokens than medium.
• Memory. Active parameters are a throughput number, not a memory number.
• Image-only benchmark charts. Several cards publish scores as pictures, which is why this page quotes few figures.

Our take

Large 4 is the interesting one: a trillion-parameter open model with a 1M window at $0.70 and $2.10, from a vendor that trains in Europe. The licence is friendlier than most large open models for application builders, and stricter for anyone who resells inference. Medium 3.5 is a transition story, and Small 4 is the clean Apache choice. We have no independent benchmark for any of them. See also open source LLM, best open source LLM for coding and GPT oss.

Product names are used descriptively. We are not affiliated with Mistral AI.