Meta releases a 30B open-weight model for on-device agents

Muse Glimmer runs at practical speeds on a single consumer GPU at 4-bit precision and leads similarly sized open models on tool use.

Meta has published the technical report for Muse Glimmer, a 30-billion-parameter open-weight model built for agentic tasks on consumer hardware. The company says it runs at practical speeds on a Mac or PC with a single consumer GPU at 4-bit precision, and that it leads comparably sized open models on tool use. The release extends the Muse family down to a size that hobbyists and small teams can actually run locally, another sign that frontier-style agent capabilities are migrating from data centers to the desk.

MiniMax ships M3.1 Flash and reveals Space Bunny

MiniMax officially launched M3.1 Flash, confirming that the previously anonymous Space Bunny, which had topped global leaderboards with more than 200 trillion tokens, was in fact an earlier version of the same model. The company frames M3.1 Flash around the performance economics of agentic AI, promising frontier-level intelligence at a fraction of the usual cost.

Prometheus, a gigawatt-scale supercomputer

Meta says its Muse family, including Muse Wonder, was trained at Prometheus in New Albany, Ohio, a supercluster with hundreds of thousands of GPUs where even one underperforming host can slow an entire training run.

FLUX 3 Max targets extreme video speed

Black Forest Labs released FLUX 3 Max, claiming it can generate a 10-second video in 12 seconds, a pace built for rapid iteration and trial-and-error workflows.

NVIDIA pushes industrial AI in Germany

In Berlin, Jensen Huang and the German chancellor discussed pairing German engineering with NVIDIA AI to advance manufacturing and the automotive industry.

BizBench stress-tests long-horizon agents

BizBench is a multi-turn benchmark built on real payment-company workflows, with 789 knowledge documents, 601 tools and 50 tasks that layer on resource constraints, time dependencies, prior commitments and conflicting guidance. Claude Opus 5 and GPT-6 Sol posted strict success rates of just 64.0% and 61.3%, while Qwen3.8-27B reached a 42.0% goal rate but slipped to 32.7% under strict checks.

AI will not automate radiologists like you think

Jason Wei pushes back on the tidy story that AI is about to replace radiologists, recalling Geoff Hinton's old claim that people should stop training radiologists now. His argument: automating a job is messier than a benchmark suggests, and the real world resists clean five-year timelines.