MENLO PARK, Calif.: Meta releases Muse Glimmer, a 30-billion-parameter open-weight model designed to run AI agents on a single consumer GPU.
Meta Superintelligence Labs announces the release on August 10. The weights ship under an unmodified Apache 2.0 license, a departure from the custom terms used for the company’s earlier Llama models. The model is available on Hugging Face, with vLLM support at launch and llama.cpp, MLX, and ExecuTorch integrations following.
The release relies on aggressive compression to fit consumer hardware. A 30-billion-parameter model requires over 55GB of memory at full precision. Meta quantized the weights to roughly 4-bit precision, reducing the model to under 20GB. Speculative decoding speeds generation approximately 3.1 times. The result runs on a 24GB graphics card or a Mac, with a bundled vision encoder for image understanding.
Muse Glimmer is distilled from Muse Spark 1.2, the closed flagship Meta launched on August 5. The model is tuned for agent workloads including function calling, local coding, and evaluation tasks. It supports more than 100 languages and can diagnose and retry failed tool calls.
Meta reports category-leading scores on MCP Atlas and SWE-Bench Pro against comparable models including Gemma4-31B and Qwen3.6-27B. The figures are the company’s own measurements and have not been independently verified.
CEO Mark Zuckerberg publishes a 14-page essay alongside the release arguing for open, distributed AI development. The release follows the launch of Meta’s first paid developer API for Muse Spark in July.
The model targets developers and enterprises running agents locally without cloud costs or data transfer. Regulated organizations can deploy it in air-gapped environments.
Previous
