Meta open-sources Muse Glimmer, a 30B local-first agentic model

Meta released Muse Glimmer today, a 30-billion-parameter multimodal model built specifically for local, agentic use — and put it out under Apache 2.0. That's the detail worth sitting with: Meta had drifted toward closed releases with its Muse line, and Glimmer is a deliberate step back toward open weights, distilled from Meta's closed Muse Spark model. It ships with day-0 support in transformers, llama.cpp, and vLLM, plus Inference Endpoints — the kind of simultaneous ecosystem support that turns a model drop into something you can actually deploy the same afternoon.

What it is

Glimmer is a dense model, not a mixture-of-experts: a 2B-parameter Perception Encoder for vision paired with a 28B-parameter text decoder. At 4-bit quantization it fits in roughly 18-20GB of VRAM, which puts it comfortably on a single consumer GPU — a 24GB card handles it with room to spare. That number is the whole thesis of the release. Meta isn't pitching Glimmer as a frontier-beating model; it's pitching it as a model you can own, self-host, and point at private documents, code, or a personal agent loop without a round trip to someone else's API.

The text decoder's attention pattern is the part I'd flag to anyone evaluating this for production. It 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 — a (SWA, SWA, SWA, Full) block repeated 13 times across 52 layers. The sliding-window layers keep local, relative structure cheap; the periodic full-attention layer, freed from position encoding, is what lets the model reason over global context without paying full quadratic cost at every layer. That's a real engineering answer to the long-context-on-a-single-GPU problem, not a marketing claim.

On top of that, grouped-query attention shares each key-value head across 16 query heads, cutting KV-cache memory by 16x — the difference between a model that chokes on long agent traces and one that doesn't. Muse Glimmer also ships an optional speculative-decoding drafter (DFlash) that Meta reports works particularly well for structured generation like code, trading some memory for materially faster output.

The vision side reuses Meta's Perception Encoder architecture — a 2B, 50-layer ViT-style tower that handles both images and video through the same pathway, patchifying frames and using pixel-shuffle to cut token count 4x without dropping channel information. Practically, that means one model, one weight file, both modalities — no bolted-on vision adapter.

How it benchmarks

Meta's published numbers put Glimmer ahead of similarly-sized open models — Gemma4-31B and Qwen3.6-27B in thinking mode — on most agentic benchmarks: MCP Atlas (75.5 vs. 54.2 and 62.5), GAIA2 (43.3 vs. 36.4 and 40.0), and DeepSearch QA. It's not uniformly ahead — Qwen3.6 edges it out on OSWorld-Verified and SWE-Bench Verified, and Gemma4 narrowly wins GPQA Diamond — but the agentic-benchmark lean tracks with how the model was built: distilled and tuned specifically for tool use and multi-step tasks rather than raw knowledge benchmarks. On safety evals, Glimmer's AgentDojo attack success rate (28.4) lands between Gemma4's better 25.6 and Qwen3.6's worse 40.3 — a reminder that smaller open agentic models are still meaningfully exploitable, and that number is worth checking before you wire one into anything with real permissions.

Why this matters

I think the interesting story here isn't the benchmark table, it's the deployment story. A model that runs agentic workloads — tool calling, multimodal reasoning, coding — locally on one GPU, under a genuinely permissive license, with same-day support across transformers, llama.cpp, and vLLM, changes the calculus for anyone building privacy-sensitive agents: document analysis on data that can't leave a machine, coding assistants that don't ship a codebase to a third party, personal-assistant setups where the cost of a cloud API call multiplied across every tool round-trip would be prohibitive. That's a meaningfully different design point than "smaller, cheaper API model," and it's the one that matters for anyone running production agent loops on their own infrastructure.

The strategic signal matters too. After a stretch of closed Muse releases, Meta shipping the distilled version openly — rather than only as an API — is a bet that open weights still buy ecosystem leverage even when the frontier model stays closed. Given the day-0 multi-library support and the benchmark positioning specifically against other open models, it reads like Meta is contesting the open-agentic-model tier on purpose, not as an afterthought.