Mistral Medium 3.5
A 128B multimodal model with a 256k window for chat, reasoning and coding
Mistral Medium 3.5 is a dense 128-billion-parameter model with a 256k context window, built by Mistral AI to handle instruction following, reasoning and coding in one set of weights. The company presents it as its first flagship merged model, replacing Mistral Medium 3.1, Magistral and Devstral 2 in its own products.
- Dense 128B parameters with a 256k context window
- Text and image input, text output
- Reasoning effort set per request, from none to high
- Company reports 77.6% on SWE-Bench Verified
- Weights under a Modified MIT License
What it is
Mistral Medium 3.5 is a large language model from Mistral AI with 128 billion parameters in a dense design, meaning every parameter is used on each step. It accepts text and pictures and replies in text. A 256k context window lets it hold very long inputs at once. The company says one set of weights covers instruction following, reasoning and programming, so users do not need separate models for those jobs. It runs in Le Chat and in the Vibe coding agent, and weights are published for anyone to download.
How it came about
Mistral AI describes this release as its first flagship merged model. It takes over from Mistral Medium 3.1 and from Magistral inside Le Chat, and it displaces Devstral 2 in the Vibe coding agent. The company says it beats its earlier coding models, the Devstral line, on every benchmark it tested. No release date is given here.
Inside the model
The model is dense rather than a mixture of experts, so all 128B parameters fire for every token. Its vision encoder was built from the ground up to cope with images of differing sizes and shapes. Reasoning effort is a per-request setting: 'none' gives a quick reply, 'high' spends extra compute at answer time for hard prompts and agent work. Suggested sampling settings are temperature 0.7 and top-p 0.95 with reasoning on, and lower temperatures with it off. An EAGLE draft model is published to speed up speculative decoding in vLLM and SGLang.
What it can do
The company reports 91.4% on τ³-Telecom and 77.6% on SWE-Bench Verified, and says it surpasses all its previous coding models on every benchmark. It handles many languages, follows system prompts closely, and supports function calling and JSON output for agent use. A configuration error in the Transformers setup hurt long-context results until a fix was committed; GGUF files built before that fix carry the same problem.
How to use it
Weights are downloadable under a Modified MIT License, which the company describes as open for commercial and non-commercial use, with carve-outs for very high revenue firms. Serving is supported through vLLM, SGLang, llama.cpp, Ollama and Transformers, with fine-tuning guides for Axolotl and Unsloth. The Mistral AI API is offered as an alternative where local hosting falls short, and the model can be picked in the Vibe CLI.
Why it matters
Within Mistral AI's own products the model consolidates several earlier ones: it takes the place of Mistral Medium 3.1, Magistral and Devstral 2, and now backs Le Chat and the Vibe coding agent. The company frames it as a single unified model for chat, reasoning and code instead of a family of specialists.
Questions people ask about Mistral Medium 3.5
What kind of model is Mistral Medium 3.5?
It is a dense 128-billion-parameter model with a 256k context window. Mistral AI built it to cover instruction following, reasoning and coding in one set of weights, and it takes text and images in, returning text.
How do I get it?
Weights are published on Hugging Face under a Modified MIT License, which Mistral AI says allows commercial and non-commercial use except for companies with very large revenue. It is also reachable through the Mistral AI API and the Vibe CLI.
What does reasoning effort do?
It is a per-request switch. Setting it to 'none' gives a fast answer, while 'high' adds test-time compute for difficult prompts and agent tasks. Mistral AI recommends 'high' for complex work and temperature 0.7 with top-p 0.95 in that mode.
Which earlier models does it replace?
Mistral AI says it supersedes Mistral Medium 3.1 and Magistral in Le Chat, and Devstral 2 in the Vibe coding agent. The company also states it outperforms all its previous coding models across the benchmarks it ran.
Can it look at pictures?
Yes. The vision encoder was trained from scratch to accept images with varying sizes and aspect ratios, and the model analyses visual content alongside text.
Good reading and sources
There is no English Wikipedia article on Mistral Medium 3.5 yet, so the facts on this page come from the sources listed above: huggingface.co/mistralai/Mistral-Medium-3.5-128B. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.