Meta's Muse Glimmer is a 30-billion-parameter open-weight model released under the Apache 2.0 license — the most permissive license Meta has ever used for an AI model. It runs locally on a single consumer GPU with 24GB of VRAM, handles text and image inputs, and scores 35 on the Artificial Analysis Intelligence Index — 21 points above Meta's last open release. Download it if you want a locally-run agent for privacy-sensitive tasks like schedule management, file organization, or coding assistance. Skip it if you need frontier-level reasoning, factual accuracy, or knowledge-work performance. Muse Glimmer is a solid local agent. It is not a replacement for Claude or GPT-6.
What Meta actually shipped
Meta released Muse Glimmer on August 10, 2026, under the Apache 2.0 license via Hugging Face. It is a distilled version of Muse Spark 1.2, Meta's proprietary flagship model, trained through a process where a smaller model learns from a larger "teacher" model.
The key specifications:
Specification | Muse Glimmer |
|---|---|
Parameters | 30B dense (~1.8B vision encoder) |
License | Apache 2.0 |
Context window | 128K default (extendable) |
Input modalities | Text + images |
Weights size (BF16) | ~60 GB |
Weights size (4-bit) | ~18 GB |
Minimum VRAM | 24GB |
Intelligence Index score | 35 |
Openness Index score | 44 |
Sources: Artificial Analysis, Meta, heise
Meta also announced plans to release the weights for Muse Spark 1.2, its more powerful flagship model, in the coming weeks.
The release was accompanied by a 6,500-word essay from Mark Zuckerberg arguing for broad distribution of AI capabilities rather than centralization among a few companies. Zuckerberg wrote: "Rather than centralising superintelligence, we should distribute it widely and give every person the ability to direct it."
Benchmark performance: strong for its size, weak where it matters
Muse Glimmer's benchmark results tell two different stories.
Where it performs well:
On agentic tool-use benchmarks, Muse Glimmer outperforms similarly sized models. It scores 75.5 on MCP-Atlas, ahead of Gemma 4 31B (54.2) and Qwen3.6-27B (62.5). On DeepSearch QA, it scores 74.6 versus Gemma 4 31B's 61.7 and Qwen3.6-27B's 71.1. On τ3-Banking, it scores 23.5 against Gemma 4 31B's 15.1.
On visual reasoning (MMMU-Pro), its vision encoder scores 74%.
Where it falls short:
Muse Glimmer scores 953 Elo on GDPval-AA v2, Artificial Analysis's leading metric for agentic knowledge work. That is below the 1,000 human baseline and behind similarly sized competitors including Qwen3.6-27B (1,141) and Gemini 3.5 Flash-Lite (1,141).
The hallucination rate is the more serious problem. Muse Glimmer's AA-Omniscience Index is -33, driven by an 82% hallucination rate compared to Qwen3.6-27B's 49%. That means when Muse Glimmer does not know something, it is far more likely to make something up than its peers.
Artificial Analysis summarized the model's profile bluntly: "Agentic knowledge work is its weakness relative to its size class."
Who should download Muse Glimmer
Privacy-sensitive users who need local processing
Muse Glimmer's primary value proposition is that it runs entirely on your hardware. No data leaves your machine. Meta designed it for "always-on local agent workflows" — schedule management, email drafting, file organization, and coding tasks that involve personal or sensitive data.
If you handle client documents, medical information, legal files, or any data you would rather not send to a cloud API, Muse Glimmer offers a genuine alternative. The tradeoff is capability: you get a 30B model instead of a frontier model, but you keep your data.
Developers building local agent workflows
Meta lists four target use cases: local agents, function calling, local coding, and LLM-as-a-judge evaluation. The model supports tool calling, multi-step task execution, file and screenshot processing, and code writing and debugging — all locally on a single GPU.
Support is broad: Ollama, LM Studio, llama.cpp, MLX, vLLM, SGLang, ExecuTorch, and Unsloth all announced integration at launch. Together AI, Fireworks AI, and OpenRouter offer hosted versions. If you are building an agent that needs to operate on local files without cloud round-trips, Muse Glimmer is the strongest open-weight option in its size class.
Commercial users who need permissive licensing
Apache 2.0 is the key differentiator from Meta's previous open models. Every prior Meta open release shipped under the Llama License, which imposed restrictions on commercial use and derivatives. Muse Glimmer has no restrictions on commercial use, modification, or redistribution.
If you want to build a product on top of an open-weight model and sell it, Muse Glimmer's license removes legal friction that Llama models carried. That alone may justify the download for some teams.
Users with 24GB+ VRAM and patience for setup
At 4-bit quantization, Muse Glimmer's weights compress to under 20GB, fitting on an RTX 3090, RTX 4090, RTX 5090, or a higher-spec MacBook. The model also supports QLoRA fine-tuning on 24GB VRAM via Unsloth, though LoRA tuning requires more than 40GB.
If you have the hardware, the barrier to entry is low. The model is free to download. You bear only your own computing costs.
Who should skip Muse Glimmer
Anyone who needs factual accuracy
An 82% hallucination rate is disqualifying for most professional work. If your task involves research, analysis, financial figures, legal citations, or any output that a human will act on, Muse Glimmer will fabricate information more often than its peers. Use a frontier model instead.
Knowledge workers who need agentic reasoning
Muse Glimmer's GDPval-AA score of 953 falls below the human baseline of 1,000. If your workflow involves realistic knowledge-work tasks — analyzing documents, synthesizing research, making judgment calls — Muse Glimmer is not ready. Qwen3.6-27B at 1,141 Elo is the better open-weight choice for this category.
Users without a capable GPU
If your machine has less than 24GB of VRAM, Muse Glimmer will not run at full context with vision enabled. The model technically supports lower-memory configurations, but performance and capability degrade. If you are on a laptop with integrated graphics or a GPU with 8–16GB VRAM, this is not the model for you.
Anyone expecting frontier-level performance
Muse Glimmer scores 35 on the Intelligence Index. Claude Opus 5.5 scores 57.6. GPT-6 Astra scores 52.7. The gap is not marginal — it is the difference between a capable local assistant and a genuine reasoning partner. If you need frontier performance, no amount of local convenience compensates.
Users who want a plug-and-play experience
Meta's download instructions require Python 3.10+, the Hugging Face CLI, and selecting the right build for your runtime from four published repositories. While Ollama and LM Studio have simplified the process, local model deployment still requires more technical comfort than opening a web app. If you want AI that just works, use a cloud provider.
How it compares to alternatives
Model | Params | License | Intelligence Index | Hallucination rate | Min VRAM |
|---|---|---|---|---|---|
Muse Glimmer | 30B | Apache 2.0 | 35 | 82% | 24GB |
Gemma 4 31B | 31B | Gemma License | 30 | Not disclosed | 24GB |
Qwen3.6-27B | 27B | Apache 2.0 | 38 | 49% | 24GB |
Kimi K2.5 | 1T total | Modified MIT | 36 | Not disclosed | Cloud |
Sources: Artificial Analysis, Hugging Face
Muse Glimmer's competitive advantage is not raw intelligence. It is the combination of Apache 2.0 licensing, a 30B size that runs on consumer hardware, and strong agentic tool-use benchmarks. Qwen3.6-27B scores higher on intelligence and hallucinates far less. Gemma 4 31B scores lower but benefits from Google's ecosystem integration.
The honest assessment: Qwen3.6-27B is the better model for most open-weight use cases. Muse Glimmer's edge is in tool calling and the permissiveness of its license. If neither of those is your priority, Qwen is the stronger choice.
FAQ
Is Muse Glimmer really free to use commercially?
Yes. Muse Glimmer is released under the Apache 2.0 license, which places no restrictions on commercial use, modification, or redistribution. This is a significant change from Meta's previous open models, which shipped under the more restrictive Llama License.
What hardware do I need to run Muse Glimmer?
A GPU with 24GB of VRAM is the practical minimum. At 4-bit quantization, the model fits on an RTX 3090, RTX 4090, RTX 5090, or a higher-spec MacBook. At full BF16 precision, you need a single H100.
Does Muse Glimmer support images?
Yes. Muse Glimmer includes a ~1.8B parameter vision encoder that scores 74% on the MMMU-Pro visual reasoning benchmark. It accepts text and image inputs, making it capable of processing screenshots, documents, and charts locally.
Why does Muse Glimmer hallucinate so much?
Meta distilled Muse Glimmer from its larger Muse Spark 1.2 model. Distillation compresses capability into a smaller model, and factual calibration is one of the capabilities that degrades most in the process. Muse Glimmer's accuracy on questions it knows is comparable to its peers, but its willingness to answer questions it does not know — rather than abstaining — is far higher.
Is Muse Glimmer better than Llama 4?
Yes, significantly. Muse Glimmer scores 35 on the Intelligence Index versus Llama 4 Maverick's 14 — a 21-point improvement. It also uses a far more permissive license and includes vision capabilities that Llama 4 Maverick lacked.
Will Meta release Muse Spark 1.2 as open weights?
Meta has announced plans to release the weights for Muse Spark 1.2, its more powerful flagship model, "in the coming weeks". Muse Spark 1.2 scores 57 on the Intelligence Index — 22 points above Muse Glimmer. If the weights are released under a similarly permissive license, it would be a far more capable open-weight model.
Should I use Muse Glimmer or a cloud API?
Use Muse Glimmer if privacy, local processing, or commercial licensing flexibility are your priorities. Use a cloud API if you need frontier-level reasoning, factual accuracy, or a plug-and-play experience. Muse Glimmer is a tool for a specific job, not a general replacement for cloud models.
Letters
No letters yet — be the first to write.