The community says
What started on X, Reddit and Hacker News before it became an announcement. 5 conversations in this edition, none confirmed yet.
Models
X
Qwen4-Max lands on Runware, You.com and other platforms
Alibaba announced that Qwen4-Max is available on providers including Runware, TokenHarbor, Lebrel and You.com.
Agents and code
X
Google Antigravity gets a robotics plugin powered by Gemini 4 Hydrogen
With Gemini 4 Hydrogen and a new robotics plugin in Google Antigravity, you can connect a compatible robotic arm, train in simulation in MuJoCo and run real-world tasks by talking in natural language.
Training and tuning
X
LoRA adapter doubles Qwen3.8-27B's performance in the pi agent
Hugging Face released Qwen3.8-27B-pi-adapted, a LoRA adapter trained with RL for the pi coding agent. With it, the SWE-bench Pro Hard score rose from 29.4% to 62.7% and Terminal-Bench 2.1 from 39.5% to 55.8%.
Reddit
ETH Zurich uses RL to teach a 5-ton loader to balance
ETH Zurich researchers used reinforcement learning to teach a 5-ton John Deere loader to balance on two wheels and cross obstacles, with no extra training on the real machine.
Coding tools
X
Mistral Vibe goes free in Europe for 3 months
Mistral made Vibe free in Europe for 3 months to celebrate the Vibe + Le Chonk pairing, which it calls the best open model-and-harness combination. The day's codes have run out.