Mistral launches Medium 3.5 for remote agents

// 142d agoMODEL RELEASE

Mistral launches Medium 3.5 for remote agents

Mistral AI has released Medium 3.5, a 128B open-weights model tuned for instruction following, reasoning, coding, and long-horizon agent workflows. The launch ties the model to Mistral Vibe remote agents and a new Work mode in Le Chat, positioning it as the execution layer for multi-step enterprise automation across code, research, inbox triage, and cross-tool workflows. Mistral says the model is in public preview, supports a 256k context window, and is now the default model in Le Chat and Vibe.

// ANALYSIS

Hot take: this is less about a model benchmark and more about Mistral shipping an opinionated agent runtime around a capable mid-size model.

–The product angle is strong: Medium 3.5 is framed as the backend for async cloud agents, not just another chat model.
–The enterprise pitch is clear: GitHub, Linear, Jira, Sentry, Slack, and Teams integration make the workflow story concrete.
–Open weights plus a modified MIT license should help adoption for teams that want self-hosting or tighter control.
–The main risk is differentiation: success depends on whether Vibe and Work mode feel meaningfully better than existing agent stacks.

// TAGS

mistralmistral-aimistral-medium-3-5agentic-airemote-agentswork-modellmopen-weightscoding-agentsenterprise-ai

DISCOVERED

142d ago

2026-04-30

PUBLISHED

142d ago

2026-04-30

RELEVANCE

9/ 10

AUTHOR

sdm1110

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

OPEN SOURCE29m ago

Chunklaya unlocks long context for Laya

chunklaya is an open-source Apache 2.0-licensed harness designed to scale Laya, an open-weights System 1 decision model, into an effective local alternative to TypeSafe's proprietary Jev model for long documents. To overcome Laya's decision-head opening bias past 200 tokens, chunklaya splits text into discrete passage chunks, replaces uncalibrated default queries with descriptive prompts to cut false positives by 25x, and uses length-sorted batching to run efficiently on consumer hardware.

OPEN SOURCE54m ago

BoundaryML drops feelings for typed AI if-statements

BoundaryML has launched feelings, an open-source project demonstrating how semantic evaluations and typed AI if-statements can be implemented natively in BAML rather than requiring a bespoke language. By pairing BAML with TypeSafe AI's Jev classifier, the library introduces methods like .feels() and .matches<T>() for fast probabilistic routing while dispatching generative tasks to LLMs via .ask().

OPEN SOURCE1h ago

Sotto debuts as an open-source, local-first macOS dictation tool powered by Whisper and Apple Silicon MLX models.

Sotto is an open-source, local-first dictation application for macOS that transcribes speech directly into active text fields upon releasing a push-to-talk hotkey. Built around a decoupled architecture, the tool pairs a native Swift menu bar client with a Bun-compiled TypeScript and Fastify inference server that orchestrates Whisper for speech recognition and MLX-accelerated language models for automatic punctuation and text cleanup. The application runs completely offline on Apple Silicon Macs while also supporting remote inference on separate Mac or Linux machines. Showcased by developer Ben Davis to explore how AI coding agents enable multi-language systems development, Sotto delivers a privacy-respecting, subscription-free alternative to proprietary voice-to-text utilities.

Mistral launches Medium 3.5 for remote agents