Muse Glimmer technical report: 30B Agent Model on One GPU
Muse Glimmer technical report landed this week: Meta details how its ~30B Apache 2.0 model runs agent tasks on a single 24GB GPU, with a drafter that hits ~233 tok/s on an RTX 5090.
- Muse Glimmer 30B is a dense model with about 29.6B parameters plus a roughly 1.8B vision encoder, 52 layers, a 131K-token context window and an Apache 2.0 license, per Meta's model card.
- Meta says 4-bit quantization keeps the language model under 20GB, so it fits on 24GB and 32GB consumer cards, with about 0.2% to 1.0% quality loss.
- On Meta's charts it leads Gemma 4 31B and Qwen 3.6 27B on agent benchmarks such as MCP Atlas, but Qwen wins several coding and computer-use tests.
in this block
The Muse Glimmer technical report is out, and it explains how Meta squeezed an agent-grade model onto a single gaming GPU. The paper, "Muse Glimmer 30B: An On-Device Agentic Model," was posted to arXiv on October 17 and promoted by @AIatMeta on October 19. The model itself is not new: its Hugging Face card lists an August 2026 release. What is new is the recipe.
What actually happened
Meta's Muse family has been busy. We covered the giant Muse Spark Open MoE and the frontier Muse Wonder release earlier. Muse Glimmer is the small sibling: distilled from Muse Spark and designed to run locally, on a PC or a high-end Mac, as the brain of a coding or browsing agent.
The weights went up in August, and the main Hugging Face repo logged 234,792 downloads in the last month, with the community GGUF builds pulling even more. The Muse Glimmer technical report adds the missing details: training stages, the distillation setup, quantization results and how Meta tested agent behavior. Meta's research page went live on October 19, which is why the model is trending again this week.
The speed trick
The headline engineering piece is a small "drafter" model for speculative decoding, which Meta calls DFlash. The drafter guesses several tokens ahead and the big model checks them in one pass. On an RTX 5090, Meta reports generation rising from 74.9 to 233.4 tokens per second, about 3.1x. On Apple silicon the gain is smaller: 23.7 to 37.8 tokens per second on an M4 Max and 26.6 to 50.2 on an M5 Max.
Those are Meta's numbers on Meta's setup. Your speed will depend on quantization, context length and which runtime you use, so treat them as a ceiling, not a promise.
The benchmarks, read honestly
Here is the comparison from the model card, Muse Glimmer versus Gemma 4 31B versus Qwen 3.6 27B:
MCP Atlas: 75.5, 54.2 and 62.5. DeepSearch QA: 74.6, 61.7 and 71.1. Those are clear wins on tool use and research-style agent tasks.
SWE-Bench Verified: 76.0, 66.6 and 77.2. OSWorld-Verified: 65.9, 58.5 and 75.6. TerminalBench 2.1: 51.7, 43.4 and 60.7. Qwen comes out ahead on all three, sometimes by a lot. Meta also lists 94.7 on AIME 2026 for math.
So Glimmer is not a universal winner in its weight class. It is a strong agent planner that is good, not best, at hands-on coding and desktop control. That is a fair tradeoff for a model built to run tools, but you should pick based on your workload.
30B parameters, one GPU, and a benchmark table that admits it loses to Qwen. Respect.
Safety and access
Meta's card says Glimmer was rated "Moderate or lower" under its preparedness framework and names agent scaffolds such as OpenClaw and Hermes Agent as tested setups. Hosted versions are available through providers including OpenRouter, Together, Featherless and DeepInfra if you do not want to run it yourself.
The Apache 2.0 license is the quiet big deal. It allows commercial use with few strings attached, including modification and redistribution, as long as you keep the license and notices. Combined with the new paper, developers now have both the weights and a written method to build on.
What to do as a reader
If you run local models, read the Muse Glimmer technical report's quantization section before downloading, and test the 4-bit build on your own agent tasks. Compare it with Qwen 3.6 27B on coding work, because the published numbers say that matchup is close.
If you just follow AI news, note that this week's buzz is about a paper, not a new launch. Be skeptical of posts calling it a fresh release, and check the model card dates yourself.
Sources: arXiv technical report and Hugging Face model card.
Not financial advice. DYOR, ser.
