Meta launches Muse Glimmer: 30B open model for always-on agents
Meta's new Muse Glimmer runs agent workflows locally with 4-bit quantization...
Meta released Muse Glimmer, a 30B open-weight model built for local agent workflows under Apache 2.0. When quantized to ~4-bit, the language model comes in under 20GB, leaving headroom in 24/32GB setups for the KV cache, perception encoder, and speculative decoding drafter. It handles interleaved text and images, supports 100+ languages, and offers controllable reasoning effort. Speculative decoding makes generation significantly faster than token-by-token with identical output quality. Trained for agentic loop tasks, it shows strong performance on DeepSearch QA, MCP-Atlas, τ3-Bench, SWE-Bench, and more—covering precise function calling, long-horizon reasoning, and failure recovery. Integrations with Ollama, LM Studio, vLLM, SGLang, and more are on the way.
- 30B open-weight model (Apache 2.0) optimized for local agent workflows with 4-bit quantization (20GB footprint)
- Supports interleaved text/images, 100+ languages, speculative decoding (2.5x faster) and multi-step agent tasks like SWE-Bench
- Available now on Hugging Face with integrations planned for Ollama, LM Studio, and optimized runtimes (llama.cpp, MLX, ExecuTorch)
Why It Matters
Enables developers to build always-on, local AI agents with minimal hardware constraints and robust failure recovery.