Published: 11 August 2026
Meta released Muse Glimmer on 10 August 2026, a 30-billion-parameter open-weights model from Meta Superintelligence Labs built for always-on local agent workflows and published on Hugging Face under the Apache 2.0 licence. Muse Glimmer is distilled from Meta’s larger Muse Spark model and compressed to roughly 4-bit precision, so it runs inside a 24 GB memory envelope on a single consumer GPU or an Apple silicon Mac. It is the first Meta agentic model to ship with open weights and a licence that permits commercial use without additional conditions.
What can Muse Glimmer do?
Muse Glimmer is trained for the agent loop rather than for open-ended chat. Meta lists end-to-end task completion, precise function calling against typed schemas, multi-step reasoning over long horizons, and failure recovery, meaning the model is trained to diagnose a failed tool call and retry rather than stop. It accepts interleaved text and images through a dedicated perception encoder, so an agent can read screenshots, charts and documents alongside the conversation. Audio is not supported, and video is processed as individual frames.
The model supports selectable reasoning strengths, so developers can trade quality against speed per request, and it works with agent scaffolds including OpenClaw. Meta trained it on data from more than 100 languages. The intended use cases are local agents, local coding tools, function-calling systems, structured extraction and LLM-as-a-judge evaluation, all of which can run without a network connection.
Muse Glimmer benchmarks and technical specs
Muse Glimmer is a dense causal transformer with roughly 30 billion total parameters, including a vision tower of about 1.8 billion parameters based on a ViT-G/14 perception encoder that accepts up to 4,096 visual tokens per image. Grouped-query attention uses 32 query heads and 2 key-value heads, and the attention stack repeats a local, local, local, global pattern with a 2,048-token sliding window. Context length is 131,072 tokens, the vocabulary is 202,048 tokens, and the knowledge cutoff is 4 January 2026.
Training ran in three phases.
- Pre-training used logit distillation on the outputs of Muse Spark, Meta’s larger foundation model.
- Mid-training added longer-context, 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 model was assessed under its Advanced AI Scaling Framework before the open-weight release and does not meet the Frontier AI definition in that framework, with chemical and biological, cyber, and loss-of-control risk rated moderate or lower. On the Siren AgentDojo prompt-injection evaluation Meta reports an attack success rate of 28.4 at a utility score of 94.2.
How does Muse Glimmer compare to Gemma4-31B and Qwen3.6-27B?
Meta positions Muse Glimmer against Google’s Gemma4-31B and Alibaba’s Qwen3.6-27B, and all comparison figures below come from Meta’s own evaluations with the comparators run in thinking mode.
- Muse Glimmer leads on MCP Atlas with 75.5 against 54.2 for Gemma4-31B and 62.5 for Qwen3.6-27B.
- It also leads on DeepSearch QA at 74.6, Gaia2 at 43.3 and SWE-Bench Pro at 51.2.
- On reasoning and instruction following Meta reports AIME 2026 at 94.7, IFBench at 77.0 and AA-LCR at 80.0.
The picture reverses on computer-use and terminal work. Qwen3.6-27B stays ahead on OSWorld-Verified with 75.6 against 65.9 for Muse Glimmer, on TerminalBench 2.1 at 60.7, and on SWE-Bench Verified at 77.2.
The consistent pattern in Meta’s own table is that Muse Glimmer wins on agentic orchestration and reasoning and trails on driving a desktop or a terminal. Meta has published a methodology report alongside the results, so the harness and settings behind each score can be checked, but no independent third party has reproduced the numbers yet.
How Muse Glimmer fits on a 24 GB consumer GPU
At full precision a 30-billion-parameter model needs more than 55 GB of memory, which is beyond any consumer GPU. Meta compresses the weights to approximately 4-bit precision, bringing the language model under 20 GB and leaving headroom for the key-value cache, the perception encoder and the speculative-decoding drafter inside a 24 GB or 32 GB envelope. Two quantised builds ship: K-Quant-Dynamic targets 32 GB of VRAM at 0.2% average accuracy degradation, and K-Quant-17GB targets 24 GB at 1.0%, both averaged across 15 common benchmarks.
Generation speed comes from DFlash, a lightweight block-diffusion drafter that proposes 16 tokens in a single forward pass, which the main model then verifies in parallel. The drafter uses 5 layers, a 2,048-token sliding window and 32 query heads against 8 key-value heads. Measuring the K-Quant-17GB build at batch size 1 with greedy decoding, Meta reports throughput rising from 74.9 to 233.4 tokens per second on an RTX 5090, a 3.1 times speedup, from 26.6 to 50.2 on an Apple M5 Max, and from 23.7 to 37.8 on an M4 Max. These are vendor-measured figures on Meta’s own harness.
Muse Glimmer availability, licence and pricing
The weights are on Hugging Face now, in a collection that carries BF16 weights, GGUF k-quants, ExecuTorch builds and the DFlash drafter, with developer documentation at Meta’s AI Developer Center. Because the licence is Apache 2.0, there is no per-token price and no usage condition beyond the licence terms: self-hosting is the day-one path, and commercial use, modification and redistribution are permitted. That is a change from Meta’s recent practice, since Muse Spark stayed closed and Muse Spark 1.1 shipped only as a hosted API.
Optimised integrations for llama.cpp, MLX and ExecuTorch are due in the coming days, with planned access through Ollama, LM Studio, Unsloth, vLLM, SGLang, Together AI, Fireworks AI and OpenRouter. Meta is working with AMD, Arm, Dell, Intel and NVIDIA on device-level optimisation. Alongside the release, Mark Zuckerberg said Meta will also open the weights for Muse Spark 1.2, its latest foundation model, although no date has been given and those weights are not yet available.
Full details, the benchmark table and the methodology report are in Meta’s announcement on Muse Glimmer, and the Muse Glimmer weights are at Huggingface.
