The August 2026 cluster pointed only at a Meta Research URL that named Muse Glimmer as an open agentic model, with no enriched facts. That gap is material for builders: a real Meta open-weight agent stack changes what can run on a laptop or a single GPU, while a teaser or a dead page would not. Primary sources now settle it. On 10 August 2026 Meta Superintelligence Labs published a research blog introducing Muse Glimmer as a 30-billion-parameter model distilled from Muse Spark, released under Apache 2.0 for always-on local agents. Independent coverage the same day from TechCrunch and Ars Technica treated the drop as a product release and as a strategic split between a smaller model people can own and more capable Muse Spark systems Meta still serves through APIs.
What we checked
- What are Muse Glimmer’s capabilities, license, weights availability, benchmarks, and intended use cases? confirmed. Muse Glimmer is a dense ~29.6B text-and-image-in, text-out model distilled from Muse Spark for local agents, coding, tool calling, multimodal reasoning, synthetic data, and LLM-as-a-judge. Weights are Apache 2.0 on Hugging Face with no gated access; Meta publishes a size-class scorecard versus Gemma4-31B and Qwen3.6-27B. [1][2][3][4]
- Is this a new model release or a research preview without usable artifacts? confirmed. It is a full open-weight release, not a preview. Meta’s 10 August 2026 blog and get-the-model docs publish ungated Hugging Face repositories for BF16, GGUF, a DFlash assistant, and ExecuTorch exports; the technical report on arXiv followed in mid-October. [1][4][7][12]
- Does Muse Glimmer outperform Gemma4-31B and Qwen3.6-27B across the board in Meta’s size-class comparison? partly confirmed. At high reasoning it leads several agent and tool scores, including MCP Atlas 75.5 versus 54.2 and 62.5, and SWE-Bench Pro 51.2 versus 36.9 and 50.2. It trails Qwen3.6-27B on OSWorld-Verified (65.9 vs 75.6), TerminalBench 2.1 (51.7 vs 60.7), SWE-Bench Verified (76.0 vs 77.2), GDPval-AA (953 vs 1141), and SkillsBench. [2][8][5]
- Did the launch-day runtime integrations actually ship, or remain a coming-days promise? confirmed. llama.cpp merged Muse Glimmer support in PR 26841 on 10 August 2026 (commit 62bf73d); official GGUFs require build b10353 or newer. Meta deploy docs now give first-class recipes for vLLM, SGLang, llama.cpp, ExecuTorch, Transformers 5.15+, Ollama 0.32.7+, LM Studio, and Unsloth. [9][7][11][10]
What we found
- Muse Glimmer is a real, ungated open-weight release. Meta’s 10 August 2026 research blog open-sources the weights under Apache 2.0. The Hugging Face model card lists authors as Meta Superintelligence Lab, a January 4, 2026 knowledge cutoff, and commercial plus research intended use. The LICENSE file on the repo is Apache License 2.0, January 2004. Get-the-model docs state there is no API key and no gated access once the files are downloaded. [1][2][3][4]
- The model is a dense causal transformer with a frozen Perception Encoder, not an image generator. The card and paper put total parameters at about 29.6B, including a ~1.8B ViT-G/14 encoder, 52 layers, hybrid local/global attention, GQA 32/2, and a 131,072-token context. Input is text plus image; output is text only. Audio is out of scope. The paper says video can be ingested as frames but is not a first-class target. Intended uses are local agents, coding agents, schema tool calling, screenshot and document reasoning, synthetic data, and LLM-as-a-judge. [2][5][6]
- Usable artifacts exist in four official Hugging Face repos: BF16 safetensors at meta-models/Muse-Glimmer-30B (~60 GB, about 72 GB VRAM to serve), GGUF K-quant 16.8 GB and dynamic 19.7 GB plus 1.4 GB mmproj and 1.6 GB DFlash drafter, a separate assistant drafter repo, and prebuilt ExecuTorch .pte exports. Meta reports 0.2% and 1.0% average accuracy degradation for the two quants versus BF16 across 15 benches. With the K-quant-17GB plus drafter, Meta measured 74.9 to 233.4 tok/s on an RTX 5090 (3.1×), 23.7 to 37.8 on M4 Max, and 26.6 to 50.2 on M5 Max. [4][7][2][5]
- Meta’s published scorecard is not a clean sweep. At high reasoning, Muse Glimmer posts MCP Atlas 75.5, DeepSearch QA 74.6, τ³-Banking 23.5, WildClawBench 47.6, SWE-Bench Pro 51.2, AIME 2026 94.7, and AA-LCR 80.0, ahead of both Gemma4-31B thinking and Qwen3.6-27B thinking on those rows. Qwen still leads OSWorld-Verified 75.6 versus 65.9, TerminalBench 2.1 60.7 versus 51.7, SWE-Bench Verified 77.2 versus 76.0, GDPval-AA 1141 versus 953, and SkillsBench 46.6 versus 44.3. The paper claims size-class state of the art on agentic tool use and planning; the table supports that only on selected rows. [2][8][5]
- Runtime support shipped on launch day rather than remaining a coming-days promise. ggml-org/llama.cpp PR 26841 merged on 10 August 2026 and is the muse-glimmer architecture used by official GGUFs from b10353 onward. Hugging Face’s launch post documents day-0 Transformers, llama.cpp, and vLLM paths. Meta’s deploy guide now covers vLLM (72 GB VRAM for BF16, parser name muse_glimmer), SGLang, llama.cpp, ExecuTorch, Transformers 5.15+, Ollama 0.32.7+ (
ollama run muse-glimmer), LM Studio, and Unsloth. Muse Glimmer is self-hosted; Meta Model API does not serve it. Together AI is documented as a hosted OpenAI-compatible option. [9][10][11][6] - Safety and product constraints are explicit. Meta says Glimmer is not Frontier AI under its Advanced AI Scaling Framework because it is generally weaker than Muse Spark, and lists chem/bio, cyber, and loss-of-control as moderate or lower (the latter two inferred from Spark 1.0). The card still reports agent-injection and privacy numbers that are not best-in-class: CI Memories violation 26.4 versus Gemma’s 12.1, Siren AgentDojo attack success 28.4 versus Gemma’s 25.6. Deployers are told to add system guardrails, confirm irreversible actions, and not offer the model to people under 18. Reasoning is on by default; llama.cpp docs say it cannot be switched off, only shortened via reasoning_strength. [2][7][11]
What it means
Treat Muse Glimmer as a local agent backbone, not as a replacement for Muse Wonder, Muse Spark, or frontier APIs. If the product needs personal context, long tool loops, and no tokens leaving the machine, Apache 2.0 plus the official GGUF and ExecuTorch builds are the point: Meta itself says this is the most permissive license it has used for an open model. Budget a 24 GB card for the 17 GB K-quant plus vision and drafter, 32 GB for the dynamic quant, and about 72 GB VRAM for BF16 serving. Use the published chat template, sample at temperature 1.0 / top_p 0.95 / top_k 64, stream long traces, never stop on <|eom|>, and pin llama.cpp at b10353 or newer. Do not read Meta’s size-class state-of-the-art claim as a blanket ranking. Qwen 3.6 27B still wins several computer-use and coding-agent rows in Meta’s own table, so run your scaffold before you standardize. Independent newsrooms correctly framed the launch as Meta letting users own a smaller agent while keeping the more capable Spark line on its API.
Still open
- Independent third-party leaderboards and harness reruns of MCP Atlas, SWE-Bench, OSWorld, and TerminalBench after the August launch remain thin relative to Meta’s internal and Artificial Analysis-sourced rows.
- How Glimmer behaves in real OpenClaw, Hermes, or coding-agent loops on consumer GPUs versus the published benches is still mostly vendor-demo evidence.
- How builders should choose among Muse Glimmer, the later Muse Spark Open line listed in Meta’s model catalog, and hosted Spark/Wonder APIs for the same agent workload is only partly specified.
- Quantization quality on long agent sessions, not the 15-benchmark average, is still a per-workload measurement.