Meta Platforms has released Muse Glimmer, a 30 billion-parameter model from Meta Superintelligence Labs, and published the weights under a permissive Apache 2.0 licence.
The weights are on Hugging Face now, with developer documentation.
Go deeper with GlobalData
Access deeper industry intelligence
Experience unmatched clarity with a single platform that combines unique data, AI, and human expertise.
Meta positions the model for always-on local agent workflows and says it runs on a Mac or PC with a single consumer GPU. The company lists local agents, function calling, local coding and LLM-as-a-judge evaluation as target use cases.
Most foundation model deployments still depend on cloud infrastructure and network access, according to Meta. It argues that local execution lets users run AI with or without an internet connection.
The company also points to open source work suggesting smaller models, trained effectively, can approach frontier-level performance on targeted tasks.
Meta describes two optimisations behind the local deployment claim. At full precision, it says, a 30 billion-parameter model would need over 55GB of memory, more than any consumer GPU offers.
Quantisation compresses the weights to roughly 4-bit precision, cutting the language model to under 20GB.
Meta says the remaining headroom holds the KV cache, the perception encoder for image understanding and a speculative decoding drafter, with all three running simultaneously within a 24GB or 32GB envelope.
The company adds that it validated the compression as causing minimal to no degradation on agentic tasks.
The release also ships a lightweight drafter based on DFlash. Meta says it proposes blocks of tokens, which the main model verifies in parallel, accepting correct tokens and correcting the rest.
The company claims this runs faster than token-by-token decoding with identical output quality. Quantised drafter versions are included to reduce memory overhead.
Meta measured throughput on a K-Quant-17GB configuration with the quantised drafter, running on MacBook M4-Max and M5-Max machines and an RTX-5090.
Meta says training ran in three phases. Pre-training used logit distillation on outputs from Muse Spark, with a data mix similar to the teacher.
Mid-training added longer-context, agent-heavy data with richer reasoning traces, alongside organic data. Post-training combined supervised fine-tuning with on-policy distillation and reinforcement learning across general, reasoning, coding and agentic domains.
The company says the model was evaluated under its Advanced AI Scaling Framework and assessed for open-weight release across all relevant categories.
Benchmarking covered DeepSearch QA, MCP-Atlas, 𝛕-Bench and SWE-Bench. Meta says Muse Glimmer performs strongly for its size class against Gemma4-31B and Qwen3.6-27B on several widely used measures.
The company lists several capabilities. It says the model handles function calls with precise schemas across extended workflows and chains reasoning over long horizons. When a tool call fails or returns an unexpected result, Meta says the model is trained to diagnose the error and retry rather than stop.
A dedicated perception encoder accepts interleaved text and images, which the company says lets agents read screenshots, charts and documents. Reasoning strength is selectable, trading quality against speed.
Meta says the model works with OpenClaw and other agentic orchestration patterns, and that training data spans more than 100 languages.
Optimised integrations for llama.cpp, MLX and ExecuTorch are due in the coming days, according to Meta. The company lists Ollama, LM Studio and Unsloth for local execution, llama.cpp, ExecuTorch and MLX for edge deployment, vLLM and SGLang for serving at scale, and Together AI, Fireworks AI and OpenRouter as hosted options.
It says further customisation is possible through PyTorch’s TorchTitan training feature, and that it is working with AMD, Arm, Dell, Intel and Nvidia on device-level performance.
The release follows Muse Code, a terminal-based coding agent built on Muse Spark 1.2. Meta has said that tool targets developers on large, complex projects, generating and verifying code while managing persistent background sub-agents across each session.
