What Is Muse Glimmer?
Muse Glimmer is a 30-billion-parameter open-weight AI model from Meta Superintelligence Labs, released on August 10, 2026. It is distilled from Muse Spark, Meta’s larger proprietary frontier model, and purpose-built for one job: running autonomous agent workflows locally, on hardware you own, without sending every prompt to a cloud API.
That last part is the whole point. Most frontier models live behind APIs. Glimmer is designed to sit inside a real agent loop on your own machine — reading screenshots, calling tools, writing and debugging code, and working through multi-step tasks from start to finish.
It arrived in the same August 10 announcement wave as Mark Zuckerberg’s “future is for everyone” post, which talked about open model weights, on-device models, and a personal superintelligence agent in the Meta AI app. Glimmer is the open-weight model from that push: the one developers can actually download and run.
Muse Glimmer Key Specifications
| Specification | Muse Glimmer |
|---|---|
| Total parameters | ~30B (29.6B), dense transformer |
| Perception encoder | ~1.8B ViT-G/14 vision encoder |
| Distilled from | Muse Spark |
| License | Apache 2.0 |
| Inputs | Text and images (video as individual frames) |
| Output | Text |
| Context window | 131,072+ tokens (128K+) |
| Vocabulary | 202,048 tokens |
| Knowledge cutoff | January 4, 2026 |
| Audio support | Not supported |
| Weight formats | BF16, GGUF k-quants, ExecuTorch builds, DFlash drafter |
| Training languages | 100+ languages |
Why Muse Glimmer Is Different
Three things set Glimmer apart from a typical open-weight release.
1. It is built for agents, not chat. Glimmer is trained and evaluated on agentic capabilities: multi-step reasoning, reliable tool and function calling, failure recovery, and working inside scaffolds such as OpenClaw and Hermes Agent. Its headline benchmarks are agent benchmarks — tool use, research, and software engineering — not generic chat scores.
2. It fits on one consumer GPU. A 30B model at full 16-bit precision needs roughly 55–60 GB of memory. Meta compresses Glimmer to roughly 4-bit quantization, bringing its footprint under 20 GB, so it runs on a single 24 GB GPU or a Mac with enough unified memory. Meta reports about 1.0% degradation from the 4-bit quantization on its own evaluations, and a better quantization tier for 32 GB cards at about 0.2% degradation. Full BF16 weights remain the option for machines with 55 GB or more.
3. It is genuinely open for commercial use. The Apache 2.0 license permits commercial use, modification, and redistribution, with an express patent grant. That is a meaningful difference from Meta’s later Muse Spark Open release, which ships under a custom Muse Model License with its own conditions. If you want to build a product on a Muse-trained model, Glimmer’s license is the simpler one.
On top of that, Meta adds DFlash, a block-level speculative decoding system (16-token blocks) that it says delivers up to a 3.1x decode speedup on an RTX 5090 — important because an agent model that answers slowly is useless inside a real-time agent loop.
Muse Glimmer Benchmarks (Meta-Reported)
All benchmark figures below are Meta’s own reported results, measured in thinking mode. Treat them as vendor claims until independent testing reproduces them.
| Benchmark | What it measures | Muse Glimmer |
|---|---|---|
| MCP Atlas | Tool use through the Model Context Protocol | 75.5 |
| DeepSearch QA | Research-style question answering | 74.6 |
| SWE-Bench Verified | Real-world software engineering tasks | 76.0 |
| SWE-Bench Pro | Harder software engineering tasks | 51.2 |
| AIME 2026 | Competition mathematics reasoning | 94.7 |
| IFBench | Instruction following | 77.0 |
| AA-LCR | Long-context reasoning | 80.0 |
The most telling comparison is MCP Atlas, where Meta publishes a full three-model table: Glimmer at 75.5, ahead of Qwen3.6-27B at 62.5 and Gemma4-31B at 54.2. Tool use through the Model Context Protocol is exactly the skill a local agent lives or dies by, so this is the number to watch if you are comparing open agent models.
Meta itself notes that Glimmer does not meet its definition of “Frontier AI” under its Advanced AI Scaling Framework, since it is generally less capable than Muse Spark. It rates chem/bio, cyber, and loss-of-control risk at moderate or lower.
How to Run Muse Glimmer Locally
Self-hosting is the day-one path. The weights are available on Hugging Face, with support in Ollama, vLLM, and LM Studio from the same week as release.
The simplest start is Ollama:
ollama run muse-glimmer
The default tag is roughly an 18 GB download — the 4-bit quantized build.
Hardware tiers, roughly:
- 24 GB GPU (or M4/M5 Max Mac): the 4-bit build, about 1% accuracy loss on Meta’s evaluations.
- 32 GB GPU: the better quantization tier, about 0.2% loss.
- 55 GB or more: full BF16 weights, no quantization loss.
- NVIDIA Blackwell cards (RTX 5090, RTX PRO 6000, B100/B200, DGX Spark) are listed as supported hardware, with vLLM as a runtime option.
Practical uses Meta targets: desktop agents that read screenshots, coding agents, schema-based function calling, multi-step planning, synthetic data generation, and evaluation of other models’ outputs. Meta advises adding system-level guardrails rather than shipping the model as a bare endpoint — sensible advice for any agent model with tool access.
Muse Glimmer vs Muse Spark vs Muse Spark Open vs Muse Code
The Muse names overlap and confuse almost everyone. Here is the clean split:
| Model / Tool | What it is | Weights | Best for |
|---|---|---|---|
| Muse Glimmer | 30B agent model, released Aug 10, 2026 | Open, Apache 2.0 | Local agents on one GPU |
| Muse Spark | Frontier model line (1.0–1.3, latest Sept 2, 2026) | Closed, API only | Top-tier performance via Meta’s API or Muse Code |
| Muse Spark Open | 816B MoE model, released Oct 15, 2026 (42B active per token, 1M context) | Open, custom Muse Model License | Frontier-class open weights on heavy hardware (hundreds of GB) |
| Muse Code | Terminal coding agent (macOS/Linux) | N/A (uses the API) | Agentic coding in your repo |
The short version: Glimmer is the Muse model you can actually host. Spark is the frontier model you rent through an API. Spark Open is the giant open model that needs a workstation-class machine with hundreds of gigabytes of memory. Glimmer does not replace Spark — Meta’s own scores put it well short of the frontier tier — but it gives you a Muse-trained agent model whose prompts stay on your server.
Who Muse Glimmer Is For (and Not For)
It is a strong fit for:
- Developers building local desktop or coding agents
- Privacy-sensitive work where code and screenshots should not leave your server
- Startups that want agent capabilities without per-token API bills
- Anyone experimenting with MCP-based tool use and agent scaffolds
It is not the right pick if:
- You need frontier-level reasoning — that is Muse Spark via API
- You need audio input — Glimmer does not support it
- You need full video understanding — video is processed as individual frames only
- You have only a basic laptop — a 24 GB GPU or high-memory Mac is the realistic floor
Limitations Worth Knowing Before You Download
- Benchmarks are vendor-reported. Meta’s numbers are promising, but independent reproduction across hardware, long contexts, and real agent workflows is still catching up.
- Quantization costs accuracy. About 1% on the 4-bit build by Meta’s measure; the exact cost depends on your task and evaluation.
- Training data details are undisclosed. Meta says 100+ languages, but dataset names and sizes are not published.
- Guardrails are your job. With tool access and local deployment, system-level safety controls are something you add, not something that ships by default.
Frequently Asked Questions
Is Muse Glimmer free?
The weights are free to download under Apache 2.0, including for commercial use. “Free” does not mean free to operate — you still need the GPU, electricity, and time to run it.
Can Muse Glimmer run on a normal laptop?
Only if it has enough memory. The realistic floor is a 24 GB GPU or a Mac with sufficient unified memory (M4/M5 Max class). A standard 16 GB laptop will not run the 30B model usefully.
Is Muse Glimmer better than Muse Spark?
No. Glimmer is distilled from Spark and is deliberately smaller and less capable. What you gain is local deployment, privacy, and no per-token API cost — not higher intelligence.
Does Muse Glimmer need the internet?
No, once downloaded it runs locally without cloud infrastructure or network access. That is one of its main selling points for private code and documents.
What is the difference between Muse Glimmer and Muse Spark Open?
Glimmer is a 30B dense model under Apache 2.0 that runs on one GPU. Spark Open is an 816B mixture-of-experts model under a custom license that needs hundreds of gigabytes of memory. Glimmer is the practical local option; Spark Open is the frontier-class open option.
Can I use Muse Glimmer commercially?
Yes. The Apache 2.0 license permits commercial use, modification, and redistribution, with an express patent grant. Standard Apache 2.0 notice requirements apply.
The Bottom Line
Muse Glimmer is Meta’s most practical open model to date: a 30B agent model, distilled from its frontier Muse Spark line, that runs on a single consumer GPU under a clean Apache 2.0 license. It will not beat the frontier models, and Meta does not claim otherwise. But for developers who want a capable local agent — one that calls tools, writes code, reads screenshots, and keeps private data on their own machine — it is currently one of the strongest open options available, and the one to watch in the local-agent space.
Sources and honesty note: Specifications and benchmark figures in this article come from Meta’s Muse Glimmer model card and release materials (via Hugging Face, NVIDIA’s model catalog, and Docker Hub listings) plus independent coverage of the August 2026 release. Benchmark scores are Meta-reported and have not been independently verified. Figures were checked against multiple sources in November 2026; if Meta updates the model card, the card itself is the authority.