Mistral releases Medium 3.5, a 128B dense open-weights model powering async cloud coding agents in Vibe and a new Work mode in Le Chat.
Key Takeaways
128B dense model, 256k context window, self-hostable on 4 GPUs; scores 77.6% on SWE-Bench Verified, ahead of Devstral 2 and Qwen3.5 397B A17B.
Open weights on Hugging Face under a modified MIT license; API priced at $1.5/M input tokens and $7.5/M output tokens.
Vibe remote agents run async in the cloud with parallel sessions; ongoing local CLI sessions can be teleported to the cloud with full state preserved.
Work mode in Le Chat adds multi-step agentic execution with connectors on by default, integrating GitHub, Linear, Jira, Sentry, and Slack.
Reasoning effort is configurable per request, so the same weights handle quick chat replies and long-horizon agentic runs without model switching.
Hacker News Comment Review
Only one comment so far: a commenter is skeptical, citing GPT-5.5 as having raised the bar high enough that other frontier releases feel underwhelming by comparison.
