Meta just open-sourced a 30B AI model that beats Google and Alibaba's best — for free

Muse Glimmer runs on a single consumer GPU and, per Meta's own numbers, beats Gemma4-31B and Qwen3.6-27B at agentic and coding tasks — and… Aliteq Lena Fischer · AI & Local Compute Editor

Muse Glimmer is a free, 30B-parameter open-weight model that beats Gemma4-31B and Qwen3.6-27B on Meta's own benchmarks and runs on a single consumer GPU.

Muse Spark 1.2 is Meta's paid, API-only reasoning model — $1.25/$4.25 per million tokens — and scores second only to Claude Opus 5 on Meta's own comparison chart.

4-bit quantization shrinks Glimmer from over 55GB to under 20GB, fitting inside a 24GB or 32GB GPU alongside its perception encoder and speculative-decoding stack.

Zuckerberg published a personal essay alongside the release, arguing AI should be 'broadly distributed' rather than controlled by a handful of institutions.

He also called for loosening restrictions on AI distillation, arguing the US 'will not be able to lead' if it keeps limiting how open models get built from closed ones.

The generous read is that Meta genuinely believes in open models. The cynical read, which I lean toward, is that Meta is the one lab that can afford to give away a 30B model and still keep its real…

Read the full story

Aliteq Meta just open-sourced a 30B AI model that beats Google and Alibaba's best — for free Read the full story on Aliteq