Muse Glimmer: a 30B open model built for always-on local agents

Meta Superintelligence Labs introduced Muse Glimmer, a 30-billion-parameter model whose weights were published under a permissive Apache 2.0 license. The company positions it for always-on local agent workflows: it is small enough, Meta says, to run on a Mac or PC with a single consumer GPU, covering local agents, function calling, local coding and LLM-as-a-judge evaluation.

What the model is built to do

Muse Glimmer was evaluated on end-to-end agentic benchmarks including DeepSearch QA, MCP-Atlas, τ-Bench and SWE-Bench, and compared with Gemma4-31B and Qwen3.6-27B in its size class. Meta says the model follows precise tool schemas across extended workflows, holds coherent plans over long horizons, and is trained to diagnose a failed tool call and retry instead of halting. A dedicated perception encoder lets it accept interleaved text and images, so an agent can read screenshots, charts and documents alongside a conversation. Meta also lists scaffold compatibility with OpenClaw, adjustable reasoning effort and training data drawn from more than 100 languages.

How it was trained

Training ran in three phases. Pre-training used logit distillation from the outputs of Muse Spark, a larger teacher model, on a similar data mix. Mid-training added longer-context, more agent-heavy data with richer reasoning traces. Post-training combined supervised fine-tuning with on-policy distillation and reinforcement learning across general, reasoning, coding and agentic domains. Meta says the release was assessed under the standards of its Advanced AI Scaling Framework.

Fitting 30B parameters onto consumer hardware

At full precision a 30-billion-parameter model needs more than 55 GB of memory, well beyond any consumer GPU. Meta quantizes the weights to roughly 4-bit precision, shrinking the language model to under 20 GB so the KV cache, the perception encoder and the speculative decoding drafter all fit inside a 24 GB or 32 GB envelope. For speed, Muse Glimmer ships with a lightweight drafter based on DFlash: the small network proposes whole blocks of tokens and the main model verifies them in parallel, correcting the wrong ones. Meta measured decode speedups of 3.1 times on an RTX 5090, 1.8 times on an M5 Max and 1.5 times on an M4 Max.

The weights are available now on Hugging Face, with developer documentation on Meta's portal. Optimized integrations for llama.cpp, MLX and ExecuTorch are promised in the coming days, alongside deployment through Ollama, LM Studio, Unsloth, vLLM, SGLang, Together AI, Fireworks AI and OpenRouter.