Meta's new local model puts a working agent on your own graphics card
Meta released Muse Glimmer on Monday: a 30-billion-parameter open-weight model built to run agents on hardware you already own, under a fully permissive Apache 2.0 licence. Squeezed down to roughly 4-bit precision, it drops under 20 GB and fits on a single high-end consumer graphics card.
Meta released Muse Glimmer on Monday: a 30-billion-parameter open-weight model built to run agents on hardware you already own, under a fully permissive Apache 2.0 licence. Squeezed down to roughly 4-bit precision, it drops under 20 GB and fits on a single high-end consumer graphics card.
What happened
Meta Superintelligence Labs published the weights on 10 August, alongside a research post framing Glimmer as an "always-on local agent" model. It's a dense model — every parameter fires on every token, no expert routing — with a vision encoder bolted on and a context window past 131,000 tokens.
The training story is the interesting part. Glimmer was distilled from Muse Spark, Meta's closed frontier model, first by copying Spark's raw output probabilities during pre-training, then with on-policy distillation and reinforcement learning after. In plain English: Meta used its big private model as a teacher, and shipped the student.
The licence matters more than the size. Apache 2.0 means you can use it commercially, modify it, and redistribute it with no acceptable-use rider attached — a real break from the Llama community licences Meta has shipped for years. Zuckerberg confirmed the release publicly and said weights for Muse Spark 1.2 will follow.
Why it matters
You can run a genuine tool-using agent offline now, and the cost is a graphics card rather than a bill that grows every month.
That changes the maths on a whole category of product. If you've been paying per token to have a model read files, click through a browser, or debug code, you've been renting a capability you can now own. For anything touching regulated or client-confidential data, "the model never leaves the building" stops being a compliance argument you have to win and becomes a deployment detail.
Getting started is already easy. Community builds landed within hours — quantised files on Hugging Face, an Apple Silicon build on Ollama, and support arriving in llama.cpp, MLX, and ExecuTorch. You do not need to wait for tooling to catch up.
The distillation angle is worth watching too. Glimmer is evidence that the gap between a lab's best closed model and the open one it teaches is narrowing to something you can measure in months rather than generations.
Key details
What it needs. Around 24 GB of graphics or unified memory at 4-bit. That's a top-end gaming card or a well-specced Mac — the ceiling of consumer hardware, not the middle of it. Call it a workstation model, not a laptop model, whatever the marketing says.
How fast. With speculative decoding — a trick where a small draft model guesses ahead and the big one checks its work in batches. Meta reports roughly a three-fold speedup on a top consumer card. Fast enough for interactive use.
Where it wins. Meta's own numbers put Glimmer ahead of similar-sized models from Google and Alibaba on agent benchmarks, tool-calling, and software engineering tasks.
Where it doesn't. It trails Qwen 3.6 on computer-use tasks and terminal work by a wide margin. All of these figures are self-reported; independent evaluations had not landed at the time of writing.
The catch nobody advertises. The memory holding your conversation competes with the model itself for the same budget. Long contexts get expensive fast on local hardware in a way they don't in the cloud. Plan your agent loops accordingly.
One more. Running locally means running without server-side moderation. Security and updates become your problem.
Bottom line
If you're building an agent on sensitive data and you have a 24 GB card sitting idle, download Glimmer this week and benchmark it against whatever API you're paying for. The licence is clean, the tooling is ready, and the worst case is you learn exactly how much cloud you actually need.
Sources
Meta AI Research — Introducing Muse Glimmer (primary, 10 August 2026) · Hugging Face — Muse-Glimmer-30B model card (primary) · NVIDIA Technical Blog · TechCrunch · MarkTechPost · Trending Topics · Unsloth GGUF builds.
Verified: release date (10 August 2026) confirmed against Meta's own research blog and Zuckerberg's post. Parameter count, Apache 2.0 licence, context length, and memory requirements cross-checked between the Meta research post, the Hugging Face model card, and NVIDIA's independent deployment guide. Flagged as self-reported: all benchmark results and speed figures come from Meta, which benchmarks Glimmer only against similar-sized open models. No independent evaluation had been published at the time of writing.