Meta releases Muse Glimmer, a 30B open-weight agent model built for one consumer GPU

Meta Superintelligence Lab released Muse Glimmer on Monday: a roughly 30-billion-parameter multimodal model distilled from Muse Spark, the closed flagship Meta introduced in April. The weights, two 4-bit quantizations, a speculative-decoding drafter and the vision encoder all ship under Apache 2.0 on Hugging Face. The design goal is narrow and unusually explicit: run an agent that calls tools, reads screenshots and edits code entirely on one consumer GPU, with no network round trip.

That framing matters more than the parameter count. Meta is not claiming a frontier model — the model card states plainly that Muse Glimmer does not meet the definition of “Frontier AI” under Meta's Advanced AI Scaling Framework, because it is generally less capable than Muse Spark. The claim is narrower: a usable agent now fits inside the memory budget of hardware developers already own.

What is in the release

• About 29.6B total parameters: a 52-layer dense text decoder plus a ~1.8B ViT-G/14 perception encoder for images.
• A context length of 131,072 tokens or more, with a knowledge cutoff of 4 January 2026.
• Text and image input, text output. Audio is unsupported, and video is handled as sampled frames — Meta says the model is not explicitly optimised for it.
• Training across more than 100 languages, with the caveat, stated in the model card, that Meta has not evaluated all of them.
• Apache 2.0 across the BF16 weights, both 4-bit variants, the DFlash drafter head and the frozen vision encoder.

The interesting engineering is memory, not scale

The decoder alternates three sliding-window layers of 2,048 tokens against one full-attention layer, repeated thirteen times to reach 52 layers. The local layers carry rotary position embeddings; the global layer carries none, which is how the model keeps relative ordering cheaply while still seeing the whole sequence periodically. Grouped-query attention shares each key-value head across 16 query heads, cutting KV-cache memory by a factor of 16. That ratio, more than any training decision, is why a 131K context is tractable on a desktop card.

Four-bit quantization brings the language model under 20 GB, leaving headroom for the KV cache, the perception encoder and the drafter inside a 24 GB or 32 GB envelope. Meta puts the accuracy cost at 0.2% for the 32 GB variant and 1.0% for the 24 GB one, averaged across fifteen benchmarks. Those are Meta's own measurements, not an outside audit.

Speed comes from DFlash, a small block-diffusion drafter that proposes sixteen tokens at once for the main model to verify in parallel. On Meta's runs at batch size one, an RTX 5090 goes from 74.9 to 233.4 tokens per second, an Apple M4 Max from 23.7 to 37.8, and an M5 Max from 26.6 to 50.2. The 3.1x on the 5090 is the headline number; the 1.5x to 1.8x on Apple silicon is the one most laptop users will actually see.

Meta's own table is not a clean sweep

The comparison set is Gemma4-31B and Qwen3.6-27B in thinking mode, and the model card notes that scores are “reported as published.” That means the columns collect numbers each vendor produced separately, rather than one harness run by one party. Read the whole table as vendor claims, including Meta's.

On agent-shaped tasks Muse Glimmer leads by wide margins: MCP Atlas 75.5 against 54.2 and 62.5, DeepSearch QA 74.6, GAIA2 43.3, SWE-Bench Pro 51.2, and long-context AA-LCR 80.0 against 68.3 and 73.3.

Qwen3.6-27B is ahead, though, in exactly the places where an agent touches a real machine: OSWorld-Verified 75.6 against Glimmer's 65.9, TerminalBench 2.1 60.7 against 51.7, SWE-Bench Verified 77.2 against 76.0, and GDPval-AA 1141 against 953. Gemma4-31B takes GPQA Diamond and Humanity's Last Exam. If your workload is computer use or terminal work, Meta's own numbers do not point at Glimmer.

The prompt-injection number deserves attention

On Siren AgentDojo, Meta reports an indirect prompt-injection attack success rate of 28.4% alongside 94.2 utility. Gemma4-31B is lower at 25.6% but with less utility; Qwen3.6-27B is worse at 40.3%. Roughly one injection in four landing is not a footnote for a model whose entire pitch is reading your files and screenshots and then acting on them. Meta's own guidance concedes the point: deploy it inside a system with additional guardrails, and require human confirmation before irreversible actions.

What changes if you build agents

• Reasoning effort is a system-prompt field rather than an API parameter — Reasoning strength: low / medium / high / xhigh — and Meta recommends high or xhigh for coding and agentic work.
• The recommended sampling configuration is temperature 1.0, top_p 0.95, top_k 64, which is looser than most people default to for tool-calling.
• It is documented as compatible with OpenClaw and Hermes Agent scaffolds, and Hugging Face says day-zero support landed in transformers, llama.cpp and vLLM.
• Meta lists LLM-as-a-judge evaluation and synthetic data generation as intended uses. Those are the two jobs where metered per-token pricing hurts most, and where moving to local weights changes the arithmetic rather than just the latency.

What we could not verify

Every performance and safety figure above originates with Meta, and no independent evaluation exists a day after release. Meta's detailed evaluation methodology sits on research.meta.ai, which was unreachable from this newsroom, as was the accompanying letter from Mark Zuckerberg — we describe that letter only as TechCrunch reported it. One more distinction worth keeping: Apache 2.0 covers the weights and released artifacts, not the training data, and Muse Spark, the model Glimmer was distilled from, remains closed.