Meta Muse Glimmer – open weights 30B local coding model
Article URL: https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model Comments URL: https://news.ycombinator.com/item?id=49241679 Points: 314 # Comments: 126
Meta Superintelligence Labs has unveiled Muse Glimmer, a 30-billion-parameter model optimized for local agent workflows. This open-source model is available under the permissive Apache 2.0 license, enabling developers to run it on their own devices, such as Macs or PCs with a single consumer GPU. Muse Glimmer is designed to support a wide range of use cases, including local agents, function calling, local coding, and LLM-as-a-judge evaluation.
It maintains strong performance on essential agentic use cases and benchmarks compared to models of similar size. By running models locally, users can access AI capabilities anytime, anywhere, without requiring an internet connection. The model's compact architecture and novel distillation technique transfer agentic reasoning from a larger teacher model while optimizing for memory and compute constraints of local hardware.
Muse Glimmer is evaluated under Meta's Advanced AI Scaling Framework, ensuring it meets the necessary criteria for open-weight release. To enhance responsiveness, the model employs quantization techniques to compress its 30-billion parameter weights to approximately 4-bit precision, reducing memory requirements to under 20 GB. This enables the model to run simultaneously with the perception encoder for image understanding and the speculative decoding drafter within a 24 GB or 32 GB memory envelope.
Additionally, Muse Glimmer features a lightweight drafter model based on DFlash, which proposes entire blocks of tokens at once, resulting in faster text generation while maintaining identical output quality. The model is available on Hugging Face, with developer documentation provided to facilitate deployment and customization through popular platforms like Ollama, LM Studio, and Unsloth.
Partnerships with AMD, Arm, Dell, Intel, and NVIDIA aim to optimize performance across various devices, while Meta's AI Developer Center offers comprehensive resources and guidance for building and deploying personal agents locally.
Written by urgent.news from Hacker News Best's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
