Mistral Medium 3.5

How it is built

Mistral's flagship is dense. All 128B parameters are used on every token.¹¹1Mistral AI, Mistral Medium 3.5 128B model card, Hugging Face: "Mistral Medium 3.5 is our first flagship merged model." https://huggingface.co/mistralai/Mistral-Medium-3.5-128B The language model stacks 88 layers with a hidden size of 12,288.²²2Mistral AI, params.json in the official repository: n_layers 88, dim 12288, n_heads 96, n_kv_heads 8, hidden_dim 28672, vocab_size 131072, moe null, qformat_weight fp8_e4m3, vision encoder with 48 layers. https://huggingface.co/mistralai/Mistral-Medium-3.5-128B/blob/main/params.json

Attention is grouped-query, with 96 query heads and 8 key-value heads in each layer.²² The feed-forward block is 28,672 wide, and the vocabulary has 131,072 tokens.²²

In front of the language model sits a vision encoder of 48 layers.²² Mistral trained it from scratch to handle variable image sizes and aspect ratios.¹¹ The published weights are FP8, in the e4m3 format.²²

What is new

Mistral calls Medium 3.5 its first flagship merged model. Reasoning and coding live in the same weights as instruction following.¹¹

It took over from Mistral Medium 3.1 and Magistral in Le Chat, and from Devstral 2 in the Vibe coding agent.¹¹

reasoning_effort takes none or high on each request. With high, Mistral recommends a temperature of 0.7. With none it allows anything from 0.0 to 0.7.¹¹

Mistral reports 77.6% on SWE-Bench Verified and 91.4% on tau3-Telecom.¹¹ The card says the model supersedes Devstral "across all benchmarks".¹¹

The API version went live on 28 April 2026, and the documentation calls it Mistral's frontier-class multimodal model.³³3Mistral AI documentation, Mistral Medium 3.5 model card, version 26.04, released 28 April 2026, "frontier-class multimodal model optimized for agentic and coding use cases", prices per million tokens, read 29 September 2026. The API release of 28 April is the earliest official date; the launch post is dated 22 May 2026. https://docs.mistral.ai/models/model-cards/mistral-medium-3-5-26-04 The launch post came on 22 May.⁴⁴4Mistral AI, launch post for Mistral Medium 3.5 and remote agents in Vibe, 22 May 2026. https://mistral.ai/news/vibe-remote-agents-mistral-medium-3-5/

Running it

The weights are under a Modified MIT License. It grants no rights to anyone whose company, or employer, had global consolidated monthly revenue above $20 million in the preceding month.⁵⁵5Mistral AI, Modified MIT License of Mistral Medium 3.5, condition 2. https://huggingface.co/mistralai/Mistral-Medium-3.5-128B/blob/main/LICENSE Those companies need a commercial licence from Mistral, or its hosted services.⁵⁵

In FP8 the weights take about 128 GB (128B x 1 byte).⁶⁶6Arithmetic: 128B parameters x 1 byte in FP8 = 128 GB for the language model weights, before the vision encoder and the KV cache. https://huggingface.co/mistralai/Mistral-Medium-3.5-128B/blob/main/params.json The launch post says self-hosting is possible "on as few as four GPUs".⁴⁴ The vLLM command on the model card uses eight.¹¹ An EAGLE draft model is published to speed up local inference.¹¹

Through Mistral's API, mistral-medium-3-5 costs $1.5 per million input tokens and $7.5 per million output tokens. Its context window is 256k.³³

The older Mistral Large 3 is much bigger, a mixture of experts of 675B released in December 2025.⁷⁷7Mistral AI documentation, Mistral Large 3 model card, version 25.12, released 2 December 2025, 675B total parameters, read 29 September 2026. https://docs.mistral.ai/models/model-cards/mistral-large-3-25-12