Meta's Superintelligence Labs has released Muse Glimmer, a 30-billion-parameter open-weight model built specifically for always-on local agent workflows. The model ships under a permissive Apache 2.0 license and is available now on Hugging Face. Meta says the design targets a practical constraint: most agentic AI still depends on cloud infrastructure, but developers increasingly need models that run on a laptop or desktop without network access.
The release arrives alongside a broader open-source push from Meta. CEO Mark Zuckerberg also confirmed that Muse Spark 1.2, the company's larger frontier model, will have its weights opened to the public. Meta's AI capital expenditure is forecast to reach $145 billion this year, and the open-weight strategy is framed as a direct response to Chinese labs such as Alibaba, DeepSeek, and Moonshot that have aggressively released competitive open models.
What's new
- Architecture: 30B dense parameters, multimodal via a dedicated perception encoder for interleaved text and images.
- Languages: Trained on data from more than 100 languages.
- Quantization: ~4-bit weight compression brings the language model under 20 GB, leaving headroom in a 24 GB or 32 GB memory envelope for the KV cache, perception encoder, and a speculative-decoding drafter.
- Speculative decoding: Ships with a lightweight DFlash-based drafter that proposes token blocks for parallel verification, delivering significantly faster generation than token-by-token decoding with identical output quality.
- Controllable effort: Adjustable reasoning strength to trade quality for speed.
- Agentic training targets: End-to-end task completion, precise function calling over long workflows, multi-step reasoning, failure recovery (diagnose and retry), and compatibility with OpenClaw and other scaffolds.
Benchmarks and capabilities
Meta's evaluation suite compares Muse Glimmer against Gemma4-31B and Qwen3.6-27B across agentic, coding, multimodal, safety, and general reasoning tasks. The model leads on several agentic benchmarks: MCP Atlas (75.5 vs. 54.2 and 62.5), DeepSearch QA (74.6 vs. 61.7 and 71.1), τ²-Banking (23.5 vs. 15.1 and 16.7), WildClawBench (47.6 vs. 37.6 and 43.2), and GAIA2 (43.3 vs. 36.4 and 40.0). Qwen3.6-27B pulls ahead on GDPval-AA (1,141 vs. 953), SkillsBench (46.6 vs. 44.3), and OSWorld-Verified (75.6 vs. 65.9).
Coding results are split. Muse Glimmer tops SWE-Bench Pro at 51.2 (Qwen 50.2, Gemma 36.9) and edges SciCode at 43.6 (Gemma 43.4, Qwen 39.8). Qwen leads SWE-Bench Verified (77.2 vs. 76.0) and TerminalBench 2.1 (60.7 vs. 51.7). On multimodal tasks, Muse Glimmer leads CharXiv Reasoning (78.8) but trails Qwen on ScreenSpot Pro (75.4 vs. 76.1) and OmniDocBench v1.5 (75.8 vs. 77.8). Safety evaluations show a CI Memories violation rate of 26.4 and a Siren AgentDojo attack success rate of 28.4, both lower than Qwen's reported figures.
Why it matters
Local agentic models change the economics of AI deployment. By fitting a capable 30B model into a 24 GB or 32 GB GPU envelope — typical of high-end consumer cards like the RTX 5090 or MacBook M4/M5 Max — Meta removes the need for cloud API calls, persistent connectivity, and per-token pricing for a wide class of personal and enterprise workflows. The Apache 2.0 license permits commercial use and modification, which matters for organizations that cannot ship data to external endpoints. The speculative-decoding drafter is a practical engineering choice: it keeps latency low enough for fluid interaction without requiring a larger model or more memory.
Our take
Muse Glimmer's real contribution is the memory budget discipline. Meta engineered the full stack — quantized weights, perception encoder, KV cache, and drafter — to coexist in 32 GB, and validated that compression does not degrade agentic tasks. That systems-level constraint is more valuable than any single benchmark lead. Developers evaluating this for production should still define their tool scaffolds and permission boundaries; the model's failure-recovery training helps, but it does not replace runtime guardrails.
