Meta has released Muse Glimmer, a 30-billion-parameter open-weights agentic model built to run on a single consumer GPU. Announced on August 10, 2026 under the permissive Apache 2.0 license, the dense model is distilled from Meta's proprietary Muse Spark and ships today on Hugging Face. Quantized to under 20 GB from a full-precision footprint of more than 55 GB, it targets a 24 GB to 32 GB VRAM envelope and has been tested on the MacBook M4-Max, M5-Max, and NVIDIA RTX 5090.
Glimmer is Meta's first open, on-device entry in the Muse family, which until now was cloud-only and proprietary. For creators and developers, that shift matters: an agent that reasons, calls tools, and reads interleaved text and images without sending a single token to an external API.
What Meta Shipped
Muse Glimmer is a dense 30B model, not a mixture-of-experts design, which keeps deployment predictable on a single card. It was trained by distillation from Muse Spark, the paid reasoning model Meta introduced earlier this summer, inheriting agentic behavior at a fraction of the size. The developer documentation lists tool use, multimodal input, and multilingual coverage as first-class capabilities rather than experimental add-ons.
Meta reports the model was trained on data from more than 100 languages and evaluated on full-task agent benchmarks including DeepSearch QA, MCP-Atlas, the tau-Bench tool-use suite, and SWE-Bench, which together measure searching, working inside scaffolds, and resolving multi-turn coding requests from start to finish. Meta positions Glimmer against Gemma4-31B and Qwen3.6-27B, the current open dense models in the same size class.
The Specs That Matter for Local Agents
The headline engineering work is compression. Meta shipped two quantization recipes, K-Quant-Dynamic and K-Quant-17GB, that squeeze the model under 20 GB with what Meta calls minimal to no degradation on agentic tasks. That is what brings a 30B agent inside the memory budget of a single RTX 5090 or an Apple Silicon laptop.
Speed comes from DFlash, a lightweight drafter model that powers speculative decoding. Meta measured a 3.1 times token-generation speedup on the RTX 5090, 1.8 times on the M5 Max, and 1.5 times on the M4 Max. Speculative decoding drafts several tokens with the small model and verifies them in one pass of the large model, so the gains are largest on the fastest hardware.
Running Muse Glimmer on Your Own Machine
Getting from download to a working local agent is a short path:
- Confirm your hardware. You need roughly 24 GB to 32 GB of VRAM or unified memory. A single RTX 5090, or an M4-Max or M5-Max Mac, is the tested target.
- Pull the weights. Download the quantized checkpoint from the meta-models/Muse-Glimmer-30B repository on Hugging Face. Pick the K-Quant-17GB build for the tightest memory fit.
- Enable DFlash. Load the drafter alongside the main model so speculative decoding is active. This is where the 1.5 to 3.1 times speedups come from.
- Wire up tools. Connect your function-calling schema or MCP servers. Glimmer is tuned for reliable tool use, so an agent loop that reads files, runs commands, and returns results works without a cloud round trip.
- Verify against a benchmark task. Run a small SWE-Bench-style fix or a multi-step search to confirm reasoning quality before you trust it on real work.
How Muse Glimmer Compares
Glimmer sits in a crowded field of open dense models near 30B. The distinction is not raw size but the combination of an open license, on-device footprint, and distillation from a frontier teacher. The table below compares it on the attributes that decide whether you can actually run it.
| Model | Parameters | License | Runs locally | Distilled from | Where to get it |
|---|---|---|---|---|---|
| Muse Glimmer | 30B dense | Apache 2.0 | Yes, single GPU under 20 GB | Muse Spark | Hugging Face |
| Muse Spark | Undisclosed | Proprietary | No, cloud API only | n/a | Meta API |
| Qwen3.6-27B | 27B dense | Open weights | Yes | n/a | Hugging Face |
| Cohere North Mini Code | 30B | Open weights | Yes | n/a | Hugging Face |
Meta's own comparison is against Qwen3.6-27B and Gemma4-31B, and it claims Glimmer performs strongly for its size class across several widely used benchmarks. If you have already tried an open 30B coder like Cohere North Mini Code, Glimmer is a direct swap to evaluate, with the added pull of Muse Spark lineage.
What This Enables for Creators
An open, on-device agent changes the economics of building with AI. There is no per-token bill, no rate limit, and no data leaving your machine, which unlocks workflows that were previously awkward to justify: batch-processing a large codebase, running an always-on background agent, or handling private client material that cannot go to a third-party API. Because Glimmer reads interleaved text and images and calls tools reliably, it can drive a local research assistant, a code-fixing loop, or a document agent entirely offline. The Apache 2.0 license means you can fine-tune it, embed it in a commercial product, and ship it without a usage clause hanging over you.
Safety and Licensing
Meta says Glimmer was evaluated under its Advanced AI Scaling Framework, the company's internal standard for testing capable models before release. The Apache 2.0 license is one of the most permissive in wide use, allowing commercial deployment, modification, and redistribution with only attribution and patent-grant terms to respect. That combination, a frontier-lineage model with a genuinely open license, is still rare among 30B releases.
What to Do Next
If you have a capable GPU or an Apple Silicon laptop, download the weights and run the DFlash-enabled build against a task you already understand so you can judge quality honestly. Compare it head to head with the open 30B model you use today, watch VRAM under load, and test your tool-calling schema before wiring it into anything that matters. The technical details, including the distillation method, are in Meta's accompanying research paper.
Frequently Asked Questions
Is Muse Glimmer really free to use? Yes. It ships under the Apache 2.0 license, which permits commercial use, modification, and redistribution. There is no per-token cost because you run the weights yourself.
What hardware do I need to run it? Meta targets a 24 GB to 32 GB VRAM or unified-memory envelope. It has been tested on the RTX 5090 and on M4-Max and M5-Max MacBooks. The quantized build is under 20 GB on disk.
How is Glimmer different from Muse Spark and Muse Code? Muse Spark is Meta's proprietary cloud reasoning model and Muse Code is its coding agent product. Glimmer is a new open-weights model distilled from Muse Spark that you can download and run locally.
What is DFlash? DFlash is a small drafter model that enables speculative decoding, giving measured speedups of 3.1 times on the RTX 5090, 1.8 times on the M5 Max, and 1.5 times on the M4 Max.
Can Muse Glimmer use tools and read images? Yes. Meta lists reliable function calling and interleaved text-and-image input as core capabilities, which is what makes it usable as a local agent rather than just a chat model.
How does it compare to other open 30B models? Meta benchmarks it against Gemma4-31B and Qwen3.6-27B and claims strong performance for its size class. The practical differentiators are the Apache 2.0 license, the sub-20 GB footprint, and its distillation from the frontier Muse Spark model.
