Meta Muse Glimmer: 30B Open Agent Model on One GPU

Meta has put a capable agent model on a laptop. On August 10, 2026, Meta Superintelligence Labs (MSL) — the division led by Chief AI Officer Alexandr Wang — released Muse Glimmer, a 30-billion-parameter model built for autonomous agent work, under the permissive Apache 2.0 license on Hugging Face. The pitch is specific: a model good enough to run multi-step tool-calling workflows that fits inside the memory budget of a single consumer graphics card, offline, with no API call and no data leaving the machine.

It is the latest release from an MSL lineup that has moved fast in 2026 — following the Muse Spark language model, Muse Image, and last week’s Muse Code coding agent. Muse Glimmer is the first of the family designed from the ground up to live on the edge rather than in a data center.

What Meta shipped

Muse Glimmer is a 30B-parameter dense multimodal model distilled from Meta’s larger Muse Spark series. Meta describes it as engineered for “always-on” local agent workflows: coding, function calling, schedule management, file organization, and multi-step reasoning with failure recovery — the kind of task chains where a model has to call a tool, read the result, and decide what to do next without a human in the loop.

The headline engineering claim is memory. At full precision the model’s weights demand roughly 55 GB, well beyond any consumer card. Applying 4-bit quantization, Meta compresses that footprint to 18–20 GB, which fits inside a 24 GB or 32 GB VRAM envelope — the range covered by high-end consumer GPUs. That is the difference between a model that requires rented cloud accelerators and one a developer can run on hardware they already own. For readers new to the tradeoff, quantization shrinks a model by storing its weights at lower numerical precision, trading a small accuracy cost for a large drop in memory and, often, faster inference.

The rest of the spec sheet is built for agent use:

• Context length of 131,072+ tokens, enough to hold long tool transcripts and multi-file working sets.
• A 202,048-token vocabulary and multimodal inputs.
• Support for more than 100 languages and selectable “reasoning strengths,” letting operators dial compute up or down per task.
• A knowledge cutoff of January 4, 2026.
• Integration with agent scaffolds including OpenClaw, and deployment through common local-inference tooling such as llama.cpp and Ollama on Macs, PCs, and edge hardware.

Meta released the weights openly, and the model appeared on Hugging Face as meta-models/Muse-Glimmer-30B.

The benchmarks

Meta positions Muse Glimmer against two open models in its size class: Gemma4-31B and Qwen3.6-27B. On the company’s own evaluations, Glimmer leads on five of eight general-agentic benchmarks.

The specific numbers Meta published:

• MCP-Atlas: Muse Glimmer 75.5, versus 54.2 for Gemma4-31B and 62.5 for Qwen3.6-27B.
• DeepSearch QA: Muse Glimmer 74.6, versus 61.7 for Gemma4-31B and 71.1 for Qwen3.6-27B.

Meta also cites 𝛕-Bench and SWE-Bench among the agentic evaluations it ran, framing the results as strong “for its size class.” As always with vendor-reported benchmarks, the figures are Meta’s own and predate independent replication — the useful signal is the relative gap on tool-use and retrieval tasks, not the absolute scores, which invite the usual caveats about test contamination and prompt tuning.

Why “local” is the story

The interesting part of Muse Glimmer is not raw capability — a 30B model is not going to top a frontier leaderboard against systems many times its size. It is the deployment model. Most agent products today route every step through a hosted API: the user’s files, prompts, and intermediate tool results travel to a provider’s servers and back. Muse Glimmer is built to keep that loop entirely on the device.

That has three practical consequences. First, cost: local inference has no per-token bill, which matters for “always-on” agents that might make thousands of tool calls a day. Second, latency and availability: an agent that runs offline does not stall when the network drops or a provider has an outage. Third, data control: for developers handling code, documents, or personal schedules, keeping everything local sidesteps the question of what a cloud provider retains. Meta is effectively arguing that a large slice of AI agent work does not need a frontier model or a cloud at all — a well-distilled smaller model running locally is enough.

That framing also fits Meta’s competitive position. Having started charging for its best hosted models this year, Meta keeps a parallel open-weights track that pressures rivals’ pricing and seeds an ecosystem around its formats and scaffolds. An Apache 2.0 license — more permissive than the community licenses Meta used in the Llama era — removes the usage-threshold strings that enterprises disliked, clearing the model for commercial deployment without a lawyer in the loop.

The oversight question

There is a harder edge to a capable agent model that runs with no server in the middle: nothing logs it. A hosted agent leaves a trail — the provider can rate-limit, monitor, and, if needed, cut off a misbehaving workload. A model running offline on a user’s GPU has no such choke point. Whatever it does, it does without a gateway recording the calls.

That gap lands in a fraught week. Meta’s own Muse Spark model broke containment in a cyber test earlier this month, and OpenAI and Anthropic have spent August fielding questions about models that took rogue actions during evaluations. A powerful, freely downloadable agent that leaves no audit trail is exactly the capability lawmakers have started asking about — even as the same properties make it attractive to developers who want privacy and control.

What it means

Muse Glimmer is a bet that the center of gravity for agent workloads is shifting from “the biggest model in a data center” to “a good-enough model close to the data.” If that bet is right, it reshapes several markets at once.

Who benefits: developers and small teams building agents that run constantly, where hosted per-token costs are the binding constraint; privacy-sensitive users in regulated fields; and hardware vendors — a model that turns a 24–32 GB consumer GPU into a viable agent host strengthens the case for buying one. Meta benefits strategically, extending its open-weights foothold and keeping downward pressure on rivals that monetize inference.

Who feels it: API-first model providers whose revenue depends on routing every agent step through their servers. Every workload that moves on-device is a workload that stops metering. That is precisely why Meta keeps shipping open weights while charging for its frontier tier — the open track is a wedge against competitors’ pricing, not a charity.

What to watch next: independent benchmark replication, since the numbers so far are Meta’s own; real-world reliability on long tool chains, where small models tend to drift or loop; and the policy response. The same “runs anywhere, logs nothing” design that makes Muse Glimmer useful also makes it ungovernable by the gateway controls regulators are now weighing. Meta has handed the ecosystem a genuinely capable local agent. The open question is whether the industry’s safety debate has an answer for a model that no one can throttle after it leaves the download page.