Meta returns to open models with Zuckerberg's plan to out-copy China and sell compute by auction

Details

After more than a year away from open-weight releases, Meta has shipped Muse Glimmer, a 30-billion-parameter model available on Hugging Face under Apache 2.0 license. It is built specifically for AI agents that need to run locally on consumer hardware around the clock: quantized to 4 bits, it fits under 20GB of memory, compatible with current consumer GPUs and MacBooks. A speculative decoding helper model accelerates text output by up to 3.1x.\n\nIn benchmark comparisons conducted by Meta, Glimmer beats Google's Gemma4-31B and Alibaba's Qwen3.6-27B on most tasks, particularly agent-focused ones like tool use, web search, and long-context reasoning. Qwen retains an edge on desktop control and terminal tasks; multimodal performance is roughly even across all three. Meta acknowledges the comparisons were run by Meta itself and may not be optimally tuned for rivals—independent third-party evals should be consulted.\n\nThe release comes alongside a Zuckerberg essay defending model distillation—Glimmer was trained by distilling Muse Spark—and calling for fewer restrictions on US AI labs. Anthropic CEO Dario Amodei pushed back, calling distillation a threat. The Wall Street Journal reports an open-weight version of Muse Spark 1.2, Meta's current strongest model, is expected in coming weeks.\n\nMeta's previous Llama 4 drew criticism for benchmark inflation and the largest variant never shipped. Whether Muse Glimmer restores credibility in the open-source community will depend on independent evaluations.