Mistral Medium 3.5

Mistral AI launched Mistral Medium 3.5 in late April 2026 as a 128B-parameter dense model with a 256k context window, now available in public preview via Mistral Vibe CLI, Le Chat, and the API. The release pairs the model with remote cloud agents in Vibe and a new Work mode in Le Chat for multi-step, cross-tool tasks, including sessions that teleport between local and cloud while preserving state. Mistral Medium 3.5 combines instruction-following, reasoning, and coding in one model, supports text and image input, and ships open weights under a modified MIT license deployable on as few as four GPUs. On coding and agentic benchmarks, Mistral Medium 3.5 reaches 77.6% on SWE-Bench Verified and is reported to outperform prior models like Devstral 2 and Qwen3.5 397B A17B. Early reactions highlighted the smooth local-to-cloud transition for autonomous coding agents and improved workflow transparency through visible agent actions. The model targets Pro, Team, and Enterprise users globally, marking a significant step in Mistral's agentic tooling roadmap.

Evals.report tracks eight benchmark results for Mistral Medium 3.5 following its April 28, 2026 release, including 77.6% on SWE-bench Verified and 94.2% pass^1 on τ²-bench Telecom, both marked Official. Other official scores include IFBench at 68.8% accuracy, GDPval at 1168 Elo, Design Arena at 1176 Elo, and FrontierCode at 8.0% weighted. The Artificial Analysis Intelligence Index entry is listed as 39.2 with an Unverified status, and AA-Omniscience sits at a notable -37. The scorecard separates official vendor-claimed results from independently verified ones, giving developers a clear provenance map for each metric. The strong τ²-bench Telecom and SWE-bench Verified numbers reinforce the model's positioning for tool-using agentic and coding workloads. AA-Omniscience's negative index value is flagged as a potential hallucination indicator worth monitoring for production deployments.

Artificial Analysis rates Mistral Medium 3.5 with an Intelligence Index of 14, placing it well above the comparable-model median of 8, while running at roughly 152 output tokens per second—faster than the 132 average. The model supports text and image input, text output, and a 256k-token context window, with total parameters listed at 128B and open weights released under a modified MIT license on Hugging Face. Pricing is set at $1.50 per 1M input tokens and $7.50 per 1M output tokens, with a 90% cache discount bringing effective cost to $0.44 per Intelligence Index task. Artificial Analysis flags the model as expensive relative to other open-weight peers of similar size, though it remains competitive on speed and conciseness. The analysis page notes this is the reasoning variant and that a non-reasoning variant may also exist.

NYU Shanghai RITS reports Mistral Medium 3.5 as a 128B dense multimodal model with a 256k context window, released April 29, 2026, under a modified MIT license with open weights on Hugging Face. It unifies instruction-following, reasoning, and coding in a single network, replacing Devstral 2 in Mistral's Vibe coding agent and supporting 24 languages, function calling, and JSON output. The model scores 77.6% on SWE-Bench Verified and 91.4% on the τ³-Telecom benchmark, with pricing at $1.5 per million input tokens and $7.5 per million output tokens. Vibe remote agents run coding sessions in isolated cloud sandboxes, integrate with GitHub, Linear, Jira, Sentry, Slack, and Teams, and can open pull requests directly.

Mistral AI released Mistral Medium 3.5 in public preview on May 22, 2026, describing it as a 128B dense model with a 256k context window that merges instruction-following, reasoning, and coding into a single set of open weights. The model ships under a modified MIT license and can be self-hosted on as few as four GPUs, with configurable reasoning effort per request. Mistral Medium 3.5 powers remote agents in Mistral Vibe and Le Chat, enabling asynchronous cloud-hosted coding sessions spawned from the CLI or chat, plus a new Work mode for multi-step research and tool workflows. API pricing is $1.5 per million input tokens and $7.5 per million output tokens, available on Pro, Team, and Enterprise plans.