Mistral Medium 3.5

Microsoft's official Source press release dated July 21, 2026 announces a significant expansion of the Microsoft–Mistral strategic partnership and explicitly names Mistral Medium 3.5 as a launch asset: "Mistral Medium 3.5 and OCR 4 are now available in Microsoft Foundry, and Mistral Medium 3.5 is now in Microsoft Copilot Studio." This makes the model directly addressable to enterprise developers building inside Microsoft's AI platform and to makers configuring agents in Copilot Studio. The release also describes a multibillion-dollar commitment from Microsoft to leverage Mistral's expanded Europe-based GPU infrastructure for AI development and delivery of Microsoft's cloud and AI services. For regulated industries, the announcement extends Microsoft's Sovereign Cloud approach by combining Mistral Medium 3.5 with Azure's ability to deploy Mistral models across cloud, cloud-connected, and fully disconnected environments while preserving data, operational, and business-continuity control. Quoted framing from Brad Smith emphasizes giving European and regulated-market customers access to frontier AI without compromising data or operational control. The integration is positioned as a distribution and deployment milestone for Mistral Medium 3.5 rather than a new model release.

BenchLM.ai's third-party profile of Mistral Medium 3.5 128B reports a capability composite score of 36.1/100 against a field median of 50.2, ranking it 130th of 194 tracked models, and lists pricing at $1.50 input / $7.50 output per million tokens with a reported 256K context window. Category breakdowns show an Agentic rank of 86/105 (percentile 18th, 3 verified benchmarks), a Coding rank of 94/135 (percentile 31st, 2 verified benchmarks), a Knowledge rank of 120/158 (percentile 24th, 2 verified benchmarks), and unmeasured Reasoning, Multimodal, Multilingual, Instruction Following, and Math categories, so the composite should be read as a partial snapshot rather than a full evaluation. The page notes that the model is most strongly evidenced for agentic workflows such as coding agents, browser research, and computer-use tasks, but it also flags that independent runtime speed has not been measured and that 7 published benchmark rows leave some tracked slots empty. Used alongside Mistral's own model card and the benchmarks.company profile, this provides an independent third-party corroboration of Mistral Medium 3.5's pricing and context, while making clear which capability dimensions still lack public evidence.

The third-party model card at benchmarks.company corroborates Mistral Medium 3.5's core technical profile, listing it as an open-weights 128B model from Mistral AI released in April 2026 with a 262K reported context length (a slight discrepancy versus Mistral's own 256K figure that should be noted) and posted pricing of $1.50 per million input tokens and $7.50 per million output tokens. Published benchmark scores include SWE-bench Verified at 77.6% resolved (agentic, developer-reported, dated 2026-04) and GPQA Diamond at 74.8% accuracy (reasoning, third-party, dated 2026-05), giving concrete reference numbers for coding-agent and reasoning evaluation. The same page provides an indicative hardware-fit table based on 128B-parameter sizing at common quantizations, showing that Mistral Medium 3.5 becomes usable roughly at 50GB of VRAM (IQ3_XXS) and comfortably fits a single high-end card like an RTX PRO 6000 (Blackwell) at 96GB or a 128GB M5 Max Mac, with FP16 weights requiring around 256GB. Independent runtime speed and additional benchmark categories (reasoning, multimodal, multilingual, instruction following, math) are not measured by this aggregator, so these figures should be read as a corroborating reference rather than a complete evaluation.

Mistral's official Hugging Face model card for Mistral Medium 3.5 128B describes it as Mistral's first flagship merged model: a dense 128B-parameter, 256k-context architecture that unifies instruction-following, reasoning, and coding in a single set of weights. It replaces Mistral Medium 3.1 and Magistral in Le Chat and replaces Devstral 2 in Mistral's coding agent Vibe, with configurable reasoning effort per request so the same weights can power both quick replies and complex agentic runs. A from-scratch vision encoder handles variable image sizes and aspect ratios for multimodal input. The card lists Mistral Medium 3.5 capabilities including a Reasoning Mode toggle, vision understanding, multilingual support across English, French, Spanish, German, Italian, Portuguese, Dutch, Chinese, Japanese, Korean, and Arabic, strong system-prompt adherence, native function calling and JSON output for agentic workflows, and a 256k context window, with weights released under a Modified MIT License. It also notes a vLLM/SGLang companion EAGLE speculative-decoding model and flags a Transformers config fix for long-context performance, urging users to regenerate affected GGUFs.

Mistral's official changelog confirms the release of Mistral Medium 3.5 (mistral-medium-3-5) on April 27, 2026, as a first-party record of the model's launch. The same page documents surrounding Mistral ecosystem updates, including OCR 4.1 going Generally Available in August 2026 and the May 2026 launch of Vibe, Mistral's unified agent that consolidates the prior Le Chat and coding experiences. Together these entries situate Mistral Medium 3.5 inside Mistral's broader product surface as the underlying flagship model for the new unified Vibe agent modes. The changelog also surfaces adjacent API and model changes relevant to developers building on Mistral Medium 3.5, such as the new include-blocks feature in the OCR API, expanded page-range syntax for OCR requests, and the June 2026 release of Leanstral 1.5 for Lean 4 formal proof engineering. It records that OCR 4 (mistral-ocr-4-0) and OCR 4.1 (mistral-ocr-4-1) became the latest OCR endpoints. While these items are not Medium 3.5 capabilities themselves, they show the model- and API-side context in which Mistral Medium 3.5 was deployed across Mistral's platform during 2026.