Meta ships its first open model since Llama 4 and floats compute sold by auction
Meta released Muse Glimmer, a 30-billion-parameter Apache 2.0 model on Hugging Face — its first open weights since Llama 4 — compressed under 20 GB to run a local agent on one consumer GPU. Meta's own benchmarks put it ahead of Google's Gemma4-31B and Alibaba's Qwen3.6-27B on most agent tasks, though Qwen still wins desktop and terminal control. Zuckerberg paired the release with an essay defending distillation of rival models and floating a 'dynamic auction' for data center capacity, the first hint of how $145 billion of spending this year turns into revenue.
Source: the-decoder.com ↗
The US shouldn't restrict open models, but make sure the best open models come from America, distillation included.
Mark Zuckerberg
Why this matters
- → Meta re-enters open models after Llama 4 stumbled, signaling renewed commitment to competitive open-source AI
- → Zuckerberg defends distilling closed models to build cheaper versions — a direct challenge to OpenAI/Anthropic
- → $600B infrastructure spend still lacks a revenue model; compute auctions are Meta's first concrete answer to i
Open-source power play