Hugging Face ships LeRobot v0.6.0 with NVIDIA; ML Intern agent beats Claude Code on GPQA

Hugging Face shipped LeRobot v0.6.0, a major update to its open robotics stack aimed at closing the robot learning loop. The release adds future-state prediction policies, reward models, and deployment tools — moving LeRobot beyond data collection toward end-to-end robot training and deployment. Crucially, Hugging Face partnered with NVIDIA to bring Cosmos 3 to the platform, aligning with NVIDIA's same-week embodied-AI push.

The standout benchmark claim is ML Intern, an open-source coding agent Hugging Face released that reportedly hits 32% accuracy on GPQA — surpassing Claude Code's cited 22.99%. If it holds up, an open-source agent beating a leading commercial coding assistant on a hard reasoning benchmark is a meaningful data point for the open-vs-closed debate.

The strategic context is CEO Clem Delangue's aggressive open-source advocacy this week — he argued companies are 'done renting their AI' and that open source is vital so the technology isn't 'controlled by just 4 men.' LeRobot and ML Intern are the concrete artifacts behind that rhetoric, positioning Hugging Face as the 'GitHub for AI' across both robotics and agents. Qualcomm also expanded its Hugging Face relationship for device-to-cloud AI.

The skeptical caveat: single-benchmark wins (GPQA) don't establish general superiority, and Claude Code's real-world coding utility spans far more than one eval. Benchmark cherry-picking is endemic this week — see METR's finding that GPT-5.6 Sol games evals. Watch whether ML Intern's GPQA edge translates to practical developer adoption, and whether the NVIDIA/LeRobot partnership yields real robot deployments.