Meta's Muse Glimmer: a 30B agent that runs on one GPU

Meta released Muse Glimmer today: a 30-billion-parameter agentic model under Apache 2.0, small enough to run on a Mac or PC with a single consumer GPU. For most small businesses the interesting number isn't the parameter count. It's that a capable local agent now fits inside hardware you can buy once, and the per-token meter stops running.

What actually happened

Meta Superintelligence Labs published Muse Glimmer on Aug 10, 2026, on Hugging Face as meta-models/Muse-Glimmer-30B, under a permissive Apache 2.0 license. Apache 2.0 matters more than the weights being public — it means commercial use, modification, and redistribution without a bespoke acceptable-use rider you have to read twice.

The engineering claim is about memory, not benchmarks. At full precision the model needs over 55 GB. Meta quantized it to roughly 4-bit precision using K-Quant techniques, dropping the language model under 20 GB and fitting the whole runtime — model, KV cache, and perception encoder — inside a 24 GB or 32 GB VRAM envelope. That is a single RTX-class card or a well-specced Mac, not a rack.

Meta positions it for local agents and function calling, local coding, and LLM-as-a-judge evaluation, and reports results on DeepSearch QA, MCP-Atlas, τ-Bench, and SWE-Bench against Gemma4-31B and Qwen3.6-27B. Those are Meta's own evaluations of Meta's own model against two competitors it chose. Treat the direction as real and the margins as marketing until someone independent reruns them. MarkTechPost describes it as derived from the larger Muse Spark family, which lines up with Meta's contributor-tier pricing play on Muse Code — the paid product on top, the open weights underneath.

Why a local AI agent matters for your business

Two things change when the model runs on your hardware.

The first is the bill. A hosted agent that reads inventory, drafts responses, and reconciles line items charges you per token, forever, at a rate the vendor sets. A local agent charges you once for the GPU and then costs electricity. For high-volume, low-stakes work — classification, extraction, first-pass drafting, tagging — that math flips fast. We've made the cost case for portable open weights before; this release moves the hardware floor down again, the same way on-device models moved it onto phones.

The second is where your data sits. If the agent reads customer records, pricing, or payroll, running it locally means those bytes never leave the building. No DPA to negotiate, no vendor retention policy to audit, no training-data opt-out to remember to click.

The honest caveat: 30B at 4-bit is not Opus or GPT-5. It will lose on long-horizon reasoning and on tasks where one wrong answer is expensive. The right architecture is usually both — local model for the volume tier, frontier API for the hard calls, with a router in between that you control. That's a build, not a subscription, which is exactly why most vendors won't sell it to you.

Sources: Meta AI Research, MarkTechPost.