Meta Ships Muse Glimmer: 30B Open-Weight Agent Model

Meta released Muse Glimmer on August 10 — a 30-billion-parameter open-weight model under the Apache 2.0 license, purpose-built for always-on local agentic workloads. This is not a general-purpose chat model. It is distilled from the Muse Spark family and explicitly designed to run autonomous agents on a single consumer GPU (24GB VRAM). Available on Hugging Face today. The weights are free — no API keys, no per-token billing.

What happened

Muse Glimmer is Meta's first open-weight model designed from the ground up for local agentic autonomy — always-on AI agents that run on your hardware, not in a cloud data center.

• 29.6B parameters, Apache 2.0 license. No commercial restrictions, no revenue-sharing clauses, no usage caps. Distilled from Muse Spark using logit distillation, mid-trained on agent-heavy data with richer reasoning traces, then post-trained with supervised fine-tuning, on-policy distillation, and reinforcement learning across general, reasoning, coding, and agentic domains.
• Consumer GPU target. At full precision the model needs 55GB+. Quantized to ~4-bit (K-Quant-17GB), it fits under 20GB — leaving room for the KV cache, the perception encoder, and a DFlash drafter, all within a 24GB VRAM envelope (RTX 3090, RTX 4090, or equivalent). A 32GB-targeted K-Quant-Dynamic variant also ships.
• 131K+ context window. Long enough for multi-hour agent sessions spanning entire codebases, documents, and tool-call chains. Trained on over 100 languages with a knowledge cutoff of January 4, 2026.
• Agent-first architecture. Designed for tool calling, multi-step reasoning, failure recovery, and persistent execution — not optimized for single-turn Q&A or creative writing. Includes a dedicated ~1.8B-parameter ViT-G/14 perception encoder for interpreting screenshots, charts, and documents alongside text.
• DFlash speculative decoding. Ships with a companion drafter model that proposes 16-token blocks in parallel, which the main model verifies. On an RTX 5090 this yields a 3.1× generation speedup (74.9 → 233.4 tok/s). On Apple M5 Max: 1.8× (26.6 → 50.2 tok/s). On M4 Max: 1.5× (23.7 → 37.8 tok/s). These are Meta's own measurements.

Meta's benchmark table compares Glimmer against Gemma4-31B and Qwen3.6-27B. Glimmer leads on agentic orchestration — MCP Atlas (75.5 vs. 54.2/62.5), DeepSearch QA (74.6), and SWE-Bench Pro (51.2 vs. 36.9/50.2). It scores 94.7 on AIME 2026. But it trails Qwen on OSWorld-Verified (65.9 vs. 75.6), TerminalBench 2.1 (51.7 vs. 60.7), and most multimodal benchmarks. All numbers are vendor-reported.

In a demo, Glimmer autonomously discovered a Home Assistant instance on the local network, queried device APIs, wrote an HTML/CSS/JavaScript dashboard from scratch, and deployed a local server to verify its own work — a multi-step agent workflow, not a chatbot Q&A.

Muse Spark 1.2 open weights are next. Mark Zuckerberg and Alexandr Wang both confirmed that open weights for Meta's frontier model, Muse Spark 1.2, will follow "in the coming weeks." If Meta follows through, it would put an actual U.S. flagship frontier model into open circulation — something no American lab has done at that tier.

Why it matters

Open-weight models that run on consumer hardware have existed since Llama. But every previous open-weight release — Llama, Mistral, Qwen, DeepSeek — was designed as a general-purpose model that could run agents. Muse Glimmer is the first open-weight model designed for agents as the primary use case.

Always-on autonomy goes local. Cloud-based agentic workflows (Claude Code, Cursor Agent, GPT-5.6) bill per token and depend on API availability. A local model running on your GPU eliminates both — your coding agent runs 24/7 without metered costs or internet dependency. Sensitive files, dev environments, and screenshots stay on your machine.

The Apache 2.0 license is a strategic shift. Meta's Llama family used a bespoke community license with a 700-million-monthly-user cap that drew years of criticism. Muse Spark is closed and metered at $4.25/M output tokens. Glimmer under Apache 2.0 — no restrictions, no usage caps, full commercial freedom — signals Meta is competing for the open-weight developer ecosystem, not just selling API access. It also gives developers (and their legal departments) unusual clarity for enterprise deployment.

The open-weight ecosystem is specializing. For two years, Chinese labs (DeepSeek, Qwen, Kimi, GLM, MiniMax) dominated open-weight releases, accounting for roughly 61% of OpenRouter token consumption by May 2026. Glimmer is a U.S.-origin counterweight — and it competes on a new axis: not raw parameter count or reasoning breadth, but agent reliability on consumer hardware.

Local agents are fundamentally unregulatable. You cannot gate a download from Hugging Face. The White House's pre-release review framework explicitly excludes open-weight models from its scope. An Apache 2.0 agentic model that runs on a home PC sits outside every current AI governance framework.

What changes for you

If you're a developer running local AI: Muse Glimmer on a 24GB GPU means you can run an always-on coding agent, research assistant, or home-automation agent locally — no API keys, no per-token billing, no internet required. Ollama 0.32.7 already supports it. LM Studio, vLLM, SGLang, and OpenRouter integrations are rolling out this week.

If you're evaluating open-weight models: Glimmer's benchmark profile is strongest on agentic orchestration and reasoning, weakest on terminal automation and OS-level tasks. Test it against Qwen3.6-27B if your workload involves shell commands or desktop GUI control; test it against Gemma4-31B if you need multimodal reasoning breadth. The DFlash speedup is worth verifying on your own hardware — the gap between 75 tok/s and 233 tok/s can make an agent feel responsive or broken.

If you're tracking AI policy: Muse Glimmer is the highest-capability open-weight agent model released to date. Combined with Meta's commitment to open-weight Muse Spark 1.2, this puts pressure on export-control and pre-release-review frameworks that have thus far avoided confronting openly distributed agentic models.

FAQ

How does Glimmer compare to Muse Spark 1.2? Glimmer is a distillation — smaller (29.6B vs. a much larger frontier model), quantized for local hardware, and open-weight under Apache 2.0. Muse Spark 1.2 remains Meta's closed-weight flagship for cloud API access at $1.25/$4.25 per 1M tokens. Glimmer sacrifices some capability for local deployment freedom. Muse Spark 1.2 open weights are promised for the coming weeks.

What hardware do I need? A single consumer GPU with 24GB VRAM (RTX 3090, RTX 4090) for the K-Quant-17GB configuration, or 32GB (RTX 5090) for the K-Quant-Dynamic variant. Apple Silicon Macs with 32GB+ unified memory (M4 Max, M5 Max) can also run the full stack. A typical 8GB or 16GB laptop cannot.

Is this a Llama replacement? Effectively yes. Meta retired the Llama brand with the April 2026 Muse Spark launch. Glimmer is the first Apache 2.0 model to replace Llama's lineage — and it carries a more permissive license than Llama ever had. Glimmer is smaller than the final Llama models but purpose-built for a specific use case, not a general-purpose drop-in.

Where are the benchmarks from? All published benchmarks are Meta's own evaluation. No independent third-party reproduction exists yet. Treat the numbers as directional — real-world agent performance on your specific tools and codebases is what matters.

What about safety? Meta assessed Glimmer under its Advanced AI Scaling Framework and determined it does not meet the "Frontier AI" threshold. Risk across chemical/biological, cyber, and loss-of-control categories was rated Moderate or lower. On Siren AgentDojo, a prompt-injection test, Glimmer shows a 28.4% attack-success rate with 94.2 utility. Meta recommends human-in-the-loop confirmation for irreversible actions.