Meta shipped Muse Glimmer this week, a 30-billion-parameter open-weight model under Apache 2.0 that is built specifically to run agentic workloads, coding, tool use, multi-turn reasoning, on a GPU you already own. That framing matters more than the parameter count does. Most agent stacks today assume a call to a hosted frontier model for every planning step, every tool decision, every retry after a failure. Muse Glimmer is Meta’s bet that a big share of that loop can move onto a single RTX 5090 or a MacBook, with no network call and no per-token bill, without giving up the multi-step reasoning that makes an agent worth running in the first place.
What is actually new here
A hybrid attention stack tuned for long local sessions. The language model alternates three sliding-window attention layers (2,048-token windows, rotary position embeddings) with one full-attention layer that uses no positional embedding at all, and repeats that four-layer block 13 times for 52 layers total. The sliding-window layers keep local token order and distance information cheap to compute; the periodic full-attention layer preserves a global view across the whole 131K-token context. It is a deliberate trade: most attention in an agent session is local (the last few tool calls, the file you are editing), so paying full quadratic cost only some of the time is the difference between a model that stays usable at 100K+ tokens of scrollback and one that grinds to a halt. A 2B ViT-style perception encoder sits in front for image and screen understanding, so the same model can look at a screenshot, read an error, and decide what to run next.
Compression and decoding speed, not just parameter count. Meta compresses the weights to roughly 4-bit with block-level speculative decoding, landing under 20GB, small enough for a single consumer GPU. The speculative decoder, called DFlash, drafts 16-token blocks and verifies them in parallel, which is where the real throughput gain comes from: 3.1x faster decoding on an RTX 5090 (74.9 to 233.4 tokens/sec in Meta’s own numbers), 1.8x on an Apple M5 Max, 1.5x on an M4 Max. Quantization alone gets you a model that fits; speculative decoding is what gets you a model that feels responsive enough to sit in an agent loop instead of a batch job.
Evaluated as an agent, not as a chatbot. Instead of leaning on static QA benchmarks, Meta reports results on end-to-end task completion suites, DeepSearch QA, MCP-Atlas, tau-Bench, and SWE-Bench, that score whether the model can actually finish a multi-turn task inside a tool-using scaffold, including recovering after something fails midway. That is the right axis to grade a model like this on. A model that answers questions well but falls apart the third time a shell command errors out is not useful as an always-on local agent, no matter how it scores on a leaderboard built for single-turn responses.
Would it run on your machine?
Quantization determines whether the model fits in memory at all; the hardware determines how fast it feels once it does. Pick what you have below.
About the numbers above
The RTX 5090, M5 Max, and M4 Max figures are Meta’s published DFlash benchmarks. Every other card’s note is a reasonable inference from the stated ~20GB 4-bit footprint, not an official Meta measurement, run your own benchmark before treating it as fact.Where this goes next
The interesting part is not that a 30B model exists, it is that Meta is explicitly optimizing for the agent loop staying on-device: long context for scrollback, tool-use and failure-recovery benchmarks instead of static QA, and a decoding path fast enough to feel interactive. If that holds up outside Meta’s own benchmark suite, it is a real option for coding agents and browser agents that need to run continuously without shipping every keystroke to a hosted API. I would not swap out a frontier model for the hardest planning steps yet, the published evals are Meta’s own and independent SWE-Bench numbers are still trickling in, but for the high-volume, low-stakes steps in an agent loop (routing, tool selection, simple edits) this is worth pulling down and testing against your own workload this week.