Meta releases Muse Glimmer, a 30B open-weight local-agent model

Meta Superintelligence Labs introduced Muse Glimmer, described as the next model from the group: a 30-billion-parameter model built for local, always-on agent workflows. The weights are open sourced under a permissive Apache 2.0 license and available now for download on Hugging Face. Meta says the model is small enough to run on a Mac or PC with a single consumer GPU, covering use cases such as local agents, function calling, local coding and use as an LLM judge for evaluation.

Meta trained Muse Glimmer in three phases. Pre-training used the outputs of Meta's larger Muse Spark model via logit distillation, drawing on a similar data mix. Mid-training added longer-context, more agent-heavy data with richer reasoning traces alongside organic data. Post-training combined supervised fine-tuning with a mix of on-policy distillation and reinforcement learning across general, reasoning, coding and agentic domains. Meta evaluated the model under its Advanced AI Scaling Framework for open-weight release.

Meta lists the capabilities the model was trained and evaluated on: end-to-end agentic task completion, measured on benchmarks including DeepSearch QA, MCP-Atlas, tau-Bench and SWE-Bench without publishing exact scores; reliable tool use with precise function-call schemas; multi-step reasoning over long horizons; failure recovery, meaning the model is trained to diagnose a failed tool call and retry rather than halt; multimodal input through a dedicated perception encoder that reads screenshots, charts and documents; compatibility with OpenClaw and other agentic orchestration scaffolds; controllable reasoning effort to trade quality for speed; and training data spanning more than 100 languages. Meta says the model compares favorably against Gemma4-31B and Qwen3.6-27B on several widely used benchmarks for its size class, though again without exact numbers.

At full precision, a 30-billion-parameter model would need over 55GB of memory, more than any consumer GPU offers. Meta compressed the weights to about 4-bit precision, shrinking the language model to under 20GB, which leaves room for the KV cache, the perception encoder and a speculative-decoding drafter model to run together inside a 24GB or 32GB memory budget. Meta says it validated that this compression causes minimal to no degradation on agentic tasks. For speed, Muse Glimmer ships with a lightweight drafter model based on DFlash that proposes blocks of tokens at once for the main model to verify in parallel, which Meta says produces identical output quality faster than standard token-by-token generation. Meta measured a quantized K-Quant-17GB variant of the model paired with the quantized DFlash drafter on a MacBook M4-Max, an M5-Max and an RTX 5090, describing the result only as fast enough for fluid conversation and real-time agent interaction, without giving tokens-per-second figures.

Key facts

Muse Glimmer is a 30-billion-parameter open-weights model from Meta Superintelligence Labs, released under an Apache 2.0 license on Hugging Face.
Quantization compresses the model from over 55GB at full precision to under 20GB at about 4-bit precision, fitting within a 24GB or 32GB memory envelope alongside its KV cache, perception encoder and speculative-decoding drafter.
A DFlash-based speculative-decoding drafter model proposes blocks of tokens that the main model verifies in parallel, which Meta says speeds generation without changing output quality.
Meta says the model compares favorably against Gemma4-31B and Qwen3.6-27B on several benchmarks for its size class, though it publishes no exact scores for benchmarks such as SWE-Bench or tau-Bench.
Training data spans more than 100 languages, and the model is built to work with tools including llama.cpp, MLX, ExecuTorch, vLLM, SGLang, Ollama and OpenRouter.

Why it matters

Muse Glimmer pushes capable agentic AI off the cloud and onto ordinary laptops and desktops, letting an agent that manages files, drafts messages or writes code run with no network connection and without sending data to a remote server. Meta frames the release as continuing its open-research tradition, this time extending open weights into agentic AI rather than just chat models, at a moment when much of the industry still routes agent workloads through hosted, frontier-scale models.

Who it affects

Developers building local-first agents and coding tools gain a free-to-use base model. Hardware makers AMD, Arm, Dell, Intel and Nvidia, and inference platforms Ollama, LM Studio, Unsloth, llama.cpp, ExecuTorch, MLX, vLLM, SGLang, Together AI, Fireworks AI and OpenRouter are named as partners for running or serving it. The release also puts pressure on the rival small open models Meta benchmarks against, Gemma4-31B and Qwen3.6-27B.

How to use it

The weights are downloadable now from Hugging Face under Apache 2.0, a license that permits commercial use. Optimized integrations with llama.cpp, MLX and ExecuTorch are due within days, and in the same window the model will become runnable locally through Ollama, LM Studio or Unsloth, servable at scale with vLLM or SGLang, and accessible through Together AI, Fireworks AI or OpenRouter; it can already be further trained with PyTorch's TorchTitan feature. Meta gives no pricing or cost figures, consistent with an open-weights release.

How solid is it

The performance claims come entirely from Meta's own announcement. It reports strong results on DeepSearch QA, MCP-Atlas, tau-Bench and SWE-Bench and says the model compares well against Gemma4-31B and Qwen3.6-27B, but it publishes no exact benchmark scores. Its speed claims, tested on a MacBook M4-Max, an M5-Max and an RTX 5090, are described only as fast enough for fluid conversation and real-time agent interaction, with no tokens-per-second numbers given. None of these figures carry independent verification.

Risks and caveats

No individual researchers or spokespeople are credited for the work, and Meta gives no release date for the promised llama.cpp, MLX and ExecuTorch integrations beyond in the coming days. The model's weights are compressed to roughly 4-bit precision; Meta says it validated minimal to no degradation from this on agentic tasks but has not published the data behind that validation. No pricing or usage costs are mentioned anywhere in the announcement.