Advertisement
Top banner ad slot
Homepage | History
Generated 2026-04-29 15:22 UTC
docs

Mistral Releases Medium 3.5 as Open Weights, Targeting Agentic and Coding Workloads

Mistral has published mistral-medium-3.5, a frontier-class multimodal model tuned for agentic and coding tasks, offered as open weights under a Modified MIT license.

Mistral Changelog <span data-utc="2026-04-29T15:21:35+00:00">2026-04-29 15:21 UTC</span> Key: evt-61a29c1f69599c8953d1 Confidence: moderate Mode: claude

Article body

Mistral released mistral-medium-3.5 on April 26, the latest entry in its flagship medium-tier line of language models. The model is described as frontier-class and multimodal, optimized specifically for agentic and coding use cases. A notable addition is a reasoning_effort parameter that allows developers to adjust the depth of the model's reasoning on the fly.

The release arrives alongside two other recent additions to the Mistral model family. Voxtral TTS (voxtral-tts-2603), released March 26, is a text-to-speech model featuring zero-shot voice cloning, multilingual support, and real-time streaming. Mistral Small 4 (mistral-small-2603), released March 16, unifies instruct, reasoning, and coding capabilities in a single multimodal package with a 256,000-token context window.

By releasing Medium 3.5 as open weights under a Modified MIT license, Mistral is signaling a push to capture developers building autonomous agents, coding assistants, and workflows where reasoning depth needs to be tuned per task. The adjustable reasoning_effort parameter is a practical lever for teams that want to trade off speed against thoroughness depending on the workload.

Why this matters

  • Mistral Medium 3.5 brings open-weights access to a frontier-class multimodal model aimed at agentic and coding tasks, directly competing with proprietary models from Anthropic, OpenAI, and Google in those domains.
  • The reasoning_effort parameter gives developers a tunable dial for reasoning depth, a feature that is rare in open-weights releases and useful for balancing latency and quality in production agent pipelines.
  • The broader model lineup — including a TTS model and a 256k-context small model — indicates Mistral is building a full-stack, open-weights alternative to closed API vendors for AI-native product teams.

Source note

  • Source: Mistral official documentation changelog at docs.mistral.ai, dated April 28. The entry confirms model name, release date, key capabilities, and license, but does not include benchmark numbers, pricing, or API availability details. Additional items (Voxtral TTS, Mistral Small 4) were cross-referenced from the same changelog feed.

Original link

Open the monitored source