Zuck rekindles open weights Llama drama with Muse Glimmer
Meta has released Muse Glimmer, a 30-billion-parameter open weights language model, marking its first such release in over a year and signalling a return to the open source AI strategy it appeared to have abandoned earlier this year. The launch comes as American AI firms face pressure over the growing dominance of Chinese open weights models, though Glimmer's relatively modest size means it is not positioned to challenge top-tier Chinese systems like Kimi K3, Qwen 3.8-Max or DeepSeek V4 Flash.
Distilled from Meta's larger proprietary Muse Spark model, Glimmer is released under the permissive Apache 2.0 licence and is aimed at smaller enterprises, local inference, coding assistants and enthusiast use, competing instead with similarly sized models from Alibaba and Google. Meta's own benchmarks suggest it outperforms Google's Gemma 4 31B and roughly matches Alibaba's Qwen 3.6-27B, though a rival Qwen 3.8-27B release is imminent. The model runs on a single high-end GPU at full precision, or on consumer cards with as little as 16GB once quantised, with speeds of 75 to 233 tokens per second reported on an Nvidia RTX 5090.
- Meta releases Muse Glimmer, its first open weights model in over a year
- 30B-parameter model targets small enterprises, not Chinese frontier rivals
- Runs on consumer GPUs; Apache 2.0 licensed for free commercial use
Both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
Advocates of Meta's open-weights approach argue that releasing models like Muse Glimmer broadens access to powerful AI beyond a handful of well-resourced firms, letting independent researchers, startups and academics inspect, fine-tune and build upon the technology rather than depending on opaque, metered APIs. They see this as a check on the concentration of AI power in a small number of companies, a spur to competition and transparency, and a way to ensure that safety research itself is not confined to the same labs building the frontier models. Some also frame it as a strategic necessity, arguing that if Western firms do not lead in open models, that space will be filled by competitors elsewhere, notably in China.
The case against
Sceptics of this strategy counter that publishing model weights removes any ability to monitor or restrict how the technology is used once released, since safeguards can be stripped out by anyone with modest technical skill. They worry this makes it easier for malicious actors to misuse increasingly capable systems for disinformation, cyberattacks or worse, with no way to revoke access after the fact, and that it hands frontier capability to authoritarian states or hostile actors without reciprocal benefit. There is also a commercial concern, namely that open-sourcing cutting-edge models allows rivals to free-ride on enormous training investments while discouraging the kind of careful, staged deployment that closed-model developers argue is necessary for responsible scaling.