Mistral Medium 3.5 Released

• Mistral AI released Mistral Medium 3.5 on 29 April 2026, announcing it alongside cloud coding agents in Vibe and a Work mode in Le Chat; it describes it as its first flagship merged model, one dense 128B-parameter network that does instruction following, reasoning and coding where Mistral previously shipped separate models^[1][2][4][5]
• It replaces Medium 3.1 and Magistral in Le Chat and Devstral 2 in the Vibe CLI, and the weights are open on Hugging Face under a modified MIT license that carves out companies with large revenue; Mistral says it can be self-hosted on as few as four GPUs^[1][4]
• Mistral's own chart puts it at 77.6 on SWE-Bench Verified, ahead of Claude Sonnet 4.5 (77.2), Kimi K2.5 (76.8) and Qwen3.5 397B (76.4) but behind Claude Sonnet 4.6 (79.6) and GLM-5.1 (80.2), and at 91.4 on tau3 Telecom against 84.9 for Sonnet 4.5 and 98.7 for GLM-5.1; the SWE-Bench figure is self-reported^[1][4]
• The picture is mixed outside coding and telecom: on tau3 Banking it scores 13.4 against 28.4 for Sonnet 4.6, and on BrowseComp 48.6 against 79.3 for GLM-5.1 and 78.6 for Qwen3.5, a Mistral run that used context management with a discard-all strategy at 100k tokens^[1]
• API pricing is $1.50 per million input tokens and $7.50 per million output tokens; the model is also hosted on NVIDIA's build.nvidia.com and offered as an NVIDIA NIM container, and reached Microsoft Foundry and Copilot Studio in July 2026^[1][2][7]
• Early reaction was split between praise as a clear step up from Devstral 2 and complaints that the price is high for the size^[3]

Notable features

• 256k-token context, text and image input, and a vision encoder trained from scratch for variable image sizes and aspect ratios^[1][4]
• Reasoning effort is chosen per request ('none' or 'high'), so one set of weights serves quick chat replies and long agentic runs; 24 languages, native function calling and JSON output^[4]
• Remote agents in Vibe run coding sessions asynchronously in isolated cloud sandboxes, in parallel, can be moved over from a local CLI session and can open GitHub pull requests; Work mode (preview) in Le Chat does multi-step tasks across connected tools with approval before sensitive actions^[1][3]
• On Mistral's math and instruction-following chart it scores 86.3 on AIME25 (avg@16), 69.0 on Allenai IFBench and 95.8 on Collie, the last ahead of every other model in the chart^[1]
• Served with vLLM (recommended), SGLang (with day-zero Docker tags), llama.cpp, Ollama and Transformers, with a companion EAGLE model for faster inference^[4]