Meta Muse Glimmer: A 30B Open-Weight Model That Runs Local Agents

Muse Glimmer is Meta's new open-weight AI model — a 30-billion-parameter, Apache 2.0 model built to run always-on AI agents on your own laptop, not in the cloud. It landed on August 10, 2026 from Meta Superintelligence Labs, distilled from the larger Muse Spark model, and it fits on a single consumer GPU. This is Meta returning to open releases with a clear bet: the future of agents is small, private, and local. Here's what Muse Glimmer is, how it benchmarks, what hardware it needs, and where it fits.

What Is Muse Glimmer?

Muse Glimmer is a 30-billion-parameter dense multimodal model from Meta Superintelligence Labs, released under a permissive Apache 2.0 license with weights on Hugging Face. Meta describes it as optimized for always-on local agent workflows — local agents, function calling, local coding, and LLM-as-a-judge evaluation — small enough to run on a Mac or PC with a single consumer GPU.

The framing matters. Meta compares Muse Glimmer against Google's Gemma and Alibaba's Qwen rather than the frontier giants, positioning it as a leader in its size class and a bet that the future isn't only enormous cloud systems but also smaller models that live on your own machine.

Meta also says open weights for Muse Spark 1.2, its stronger foundation model, are coming next — so Muse Glimmer is likely the first of a new open-weight family, not a one-off. (CNBC)

Specs at a Glance

A few of these deserve unpacking below — especially the benchmarks and the local-hardware story, which is the whole point of the release.

How Meta Trained It

Muse Glimmer is a distillation play: it packs a bigger model's know-how into something small enough to run at home. Per Meta's model card, training ran in three phases:

• Pre-training on Muse Spark's outputs using logit distillation, with a similar data mix to the teacher model.
• Mid-training on longer-context, more agent-heavy data with richer reasoning traces, alongside organic data.
• Post-training combining supervised fine-tuning with on-policy distillation and reinforcement learning across general, reasoning, coding, and agentic domains. (Meta AI)

Meta says it evaluated the model under its Advanced AI Scaling Framework and assessed it for open-weight release across all relevant safety categories.

What Muse Glimmer Is Built to Do

Meta pitches this as a genuine agent rather than a chatbot. The capabilities it highlights map directly onto real agent workflows:

• End-to-end task completion across benchmarks like DeepSearch QA, MCP-Atlas, 𝜏-Bench, and SWE-Bench — working inside scaffolds, writing and debugging code, and resolving multi-turn requests start to finish.
• Reliable tool use — invoking functions with precise schemas throughout long workflows.
• Multi-step reasoning over long horizons, sustaining coherent plans across extended tasks.
• Failure recovery — when a tool call fails or returns something unexpected, the model is trained to diagnose and retry rather than halt.
• Multimodal input via a ~1.8B perception encoder, so agents can interpret screenshots, charts, and documents alongside conversation.
• Controllable effort — a dial to trade reasoning depth against speed.

Community testers have flagged tool-calling in particular as a standout for its size — the capability that matters most when a model is the engine behind an autonomous agent.

Benchmarks: How It Stacks Up

Meta positions Muse Glimmer against two same-class open models — Gemma4-31B and Qwen3.6-27B — and reports strong agentic numbers for a 30B model. Coverage of the launch notes it outperforms both across benchmarks like MCP Atlas and SWE-Bench Pro. Claimed results include:

Two honest caveats. First, these are vendor-reported numbers on the day of launch — Meta has a history here, having previously conceded it used specialized, unreleased variants to boost some Llama 4 benchmark scores. Treat the table as a starting point, not a verdict. Second, results against Qwen are mixed on some coding and desktop-agent benches; Muse Glimmer isn't a clean sweep. The fair read: it's competitive-to-leading in its size class on agentic tasks, not a frontier-model killer.

Running Muse Glimmer Locally

This is where the release earns its "local agent" pitch. At full precision, a 30B model needs over 55 GB of memory — more than any consumer GPU. Meta's official ~4-bit quantization compresses the language model to under 20 GB, leaving headroom for the KV cache, the perception encoder, and the speculative-decoding drafter inside a 24 GB or 32 GB envelope. Meta says the compression introduces minimal to no degradation on agentic tasks.

That means it runs on ordinary hardware — a MacBook or a single RTX 5090 — with day-one support across local-AI tools like Ollama, LM Studio, and vLLM.

To keep it feeling responsive, Muse Glimmer ships with a lightweight DFlash "drafter" — a small companion network that proposes whole blocks of tokens at once, which the main model verifies in parallel. Meta reports significant speedups from this speculative decoding while producing identical output quality, measured on M4-Max and M5-Max MacBooks and an RTX 5090.

Deployment reach is broad: llama.cpp, ExecuTorch, and MLX for edge; vLLM and SGLang to serve at scale; Together AI, Fireworks AI, and OpenRouter for hosted access; and PyTorch's TorchTitan if you want to fine-tune. (Meta AI)

The Politics Behind the Launch

Muse Glimmer isn't only a product — it's a lobbying prop. Mark Zuckerberg paired the launch with an essay arguing the U.S. must lead the open-source AI ecosystem, and that American labs face extra friction — chiefly restrictions on training data — that foreign rivals don't. His prescription: level the playing field at home rather than ban foreign models, which he dismissed as ineffective. He openly championed distillation, the very technique that produced Muse Glimmer.

The subtext is a race Meta worries about losing. Chinese developers now set the pace in open-weight AI, with models like Moonshot's Kimi K3 and Alibaba's Qwen leading, and DeepSeek releasing some of the cheapest capable models to run. Muse Glimmer is Meta's answer — an efficient, downloadable system in the same mold.

There's also an investor angle. Meta is trying to reassure the market that its enormous AI spend — capex forecast up to $145 billion this year — is paying off, and that Superintelligence Labs, formed last year under Alexandr Wang, is making progress against OpenAI and Anthropic. Open weights are the part of that strategy built to win developers and goodwill rather than direct revenue.

It's a notable reversal. Meta had appeared to go proprietary with its recent Muse Spark phase to chase closed rivals; with inference costs surging and enterprises wanting control, it's back on the open bandwagon.

Bottom Line

Muse Glimmer is today's headline: a 30B Apache-2.0 open-weight local-agent model you can download now, with Muse Spark 1.2 open weights promised next. It's less about beating GPT or Claude at frontier scale and more about strong, private, on-device agents — reliable tool use and multi-step reasoning that fit on one consumer GPU. If your roadmap includes agents that run locally and keep data on your own hardware, it belongs on the shortlist alongside Gemma and Qwen.

Put an Open-Weight Model Like This to Work Locally

A model like Muse Glimmer only matters once it's wired into real work — tools, files, code, and multi-step plans — not a chat box. That's what a model-agnostic, multi-agent platform is for. Eigent is an open-source Cowork desktop app that runs an AI workforce locally, so you can plug in the best open-weight model per task and keep sensitive data on your own machine. See how agents can review GitHub PRs end to end, then download Eigent to try it.

Frequently Asked Questions

What is Meta Muse Glimmer?

Muse Glimmer is a 30-billion-parameter, open-weight multimodal model from Meta Superintelligence Labs, released August 10, 2026 under an Apache 2.0 license. It's distilled from Meta's larger Muse Spark model and optimized for always-on local AI agents that run on a single consumer GPU.

Is Muse Glimmer really open source?

Meta released the weights on Hugging Face under the permissive Apache 2.0 license, which allows downloading, modifying, self-hosting, and commercial use. That's more permissive than the older Llama community licenses. Meta has also said open weights for its stronger Muse Spark 1.2 model are coming next.

What hardware do I need to run Muse Glimmer?

At full BF16 precision the model needs roughly 55–60 GB of memory. Meta's official ~4-bit quantization shrinks it to under 20 GB, so it fits within a 24 GB or 32 GB envelope — a single high-end consumer GPU like an RTX 5090, or a capable MacBook. It runs day one on Ollama, LM Studio, and vLLM.

How does Muse Glimmer compare to Gemma and Qwen?

Meta benchmarks it against Gemma4-31B and Qwen3.6-27B and reports strong agentic results for its size — for example 75.5 on MCP Atlas versus 54.2 and 62.5. Results against Qwen are mixed on some coding and desktop-agent tasks, and all numbers are vendor-reported at launch, so treat them as a starting point.

What is Muse Spark 1.2?

Muse Spark 1.2 is Meta's stronger foundation model — the larger sibling Muse Glimmer is distilled from. Zuckerberg said Meta will open the weights for Muse Spark 1.2, extending its return to open releases beyond the smaller Glimmer model.

Can I use Muse Glimmer with a multi-agent platform?

Yes. Because it's open-weight and self-hostable with day-one support in local runtimes, a model-agnostic platform like Eigent can route agent tasks to Muse Glimmer while keeping data on your own machine — useful when privacy or on-device operation matters.