Meta Went Apache 2.0 on a 30B Agent Model. The Flagship Is Next.

TL;DR

Meta released Muse Glimmer this morning: a 30B-parameter, dense, multimodal model tuned for local agent workloads, with weights on Hugging Face under a straight Apache 2.0 license. It posts the best agentic-benchmark numbers in its size class, runs quantized in 24 GB of VRAM, and ships with day-one support across the local-inference ecosystem. The bigger headline is the promise attached: open weights for a version of Muse Spark 1.2, Meta's flagship, in the coming weeks. After a month of MIT licenses with tripwires and open models Europeans are forbidden to run, an actual Apache 2.0 reads like a typo.

What shipped

Muse Glimmer is a 29.6B dense causal transformer (52 layers, 32 query heads with 2 KV heads, RoPE, a repeating local-local-local-global attention pattern with a 2,048-token sliding window) paired with a 1.8B ViT-style perception encoder for image input. Context is 131,072 tokens, the vocabulary is 202,048, training covered 100+ languages, and the knowledge cutoff is January 4, 2026. Text and images go in; text comes out. No audio in either direction.

The launch is unusually well-coordinated for a weights drop: Ollama, LM Studio, Unsloth, vLLM, SGLang, llama.cpp, MLX, plus hosted endpoints at Together AI, Fireworks AI, and OpenRouter, all on day one. This is a model Meta wants running on your machine tonight, not admired from a leaderboard.

A specialist, and honest about it

Meta pitches Glimmer at "always-on" local agents: tool use, long multi-step tasks, failure recovery. The published numbers back the positioning. Against Gemma4-31B and Qwen3.6-27B, the two obvious rivals in the 30B weight class, Glimmer opens a wide gap on agentic evals: 75.5 on MCP Atlas (vs 54.2 and 62.5), 74.6 on DeepSearch QA (vs 61.7 and 71.1), 47.6 on WildClawBench, 43.3 on Gaia2.

Outside the agent lane it is merely competitive, and the model card does not hide it. SWE-Bench Verified lands at 76.0, a hair under Qwen3.6-27B's 77.2, though Glimmer flips the harder SWE-Bench Pro (51.2 vs 50.2). GPQA Diamond is 83.5 against Gemma4's 85.7. AIME 2026 is a strong 94.7. Read the table the way Meta clearly wants you to: this is not a frontier all-rounder, it is a purpose-built local agent brain that happens to be decent at everything else.

Will it run on your box?

The full BF16 weights want 64 GB. Two official quants change the math: K-Quant-Dynamic targets 32 GB with a claimed 0.2% quality loss, and K-Quant-17GB fits a 24 GB card at a claimed 1.0%. That puts the sweet spot exactly on an RTX 5090, a 4090 with the smaller quant and some KV-cache discipline, or any recent Mac with unified memory to spare.

Speed comes from DFlash, Meta's speculative-decoding setup. The mechanism: a small draft model proposes several tokens ahead, and the big model verifies the whole batch in a single forward pass, which is far cheaper than generating each token itself. Think of a fast intern drafting the next few sentences while the senior partner only nods or strikes them out; approving is cheaper than writing. On the quantized model that yields 233.4 tok/s on an RTX 5090 (a 3.1x speedup), 50.2 tok/s on an M5 Max, and 37.8 tok/s on an M4 Max.

The bigger flip

Glimmer alone would be a good local-AI story. The strategic news is what Meta attached to it: the Wall Street Journal reports Meta will release an open-weight version of Muse Spark 1.2, its most advanced model, in the coming weeks. Meta superintelligence chief Alexandr Wang confirmed it: "two very different models, both headed into people's hands, with more to come." Until now the Muse era has been API-only, and Meta spent last week selling Spark 1.2 access at 95 percent off in exchange for training rights on your code. Turns out there was a lower price point.

Mark Zuckerberg supplied the ideology the same morning in a long essay, "The Future Is For Everyone," arguing that "open source is a positive and important force for empowering people and preventing centralization," that Meta "will resume releasing some open source models soon," and that "there is no such thing as a singular benevolent superintelligence." The essay lands squarely in the open-weights policy fight of the past two weeks, on the side of the 50-signer open-weights letter and against restriction-first proposals.

Caveats before you clear disk space

All benchmark and quant-degradation numbers above are Meta's own; no third-party evals existed at publish time. The Spark 1.2 commitment is a pledge, not a release, and "an open-weight version of" is doing quiet work in that sentence: nobody has said the open weights will match the hosted flagship. Audio input and output are unsupported, and the model card recommends guardrails before wiring Glimmer into agentic deployments that can take irreversible actions.

Key Takeaways

• Meta released Muse Glimmer, a 29.6B dense multimodal model with a 131K context, under Apache 2.0, with weights live on Hugging Face and day-one support in Ollama, LM Studio, vLLM, llama.cpp, and MLX.
• It leads its size class on Meta's published agentic evals (75.5 MCP Atlas vs 62.5 for Qwen3.6-27B and 54.2 for Gemma4-31B) while staying merely competitive on GPQA and SWE-Bench Verified.
• Official quants fit 32 GB and 24 GB envelopes; DFlash speculative decoding hits 233.4 tok/s on an RTX 5090.
• Meta pledged open weights for a version of flagship Muse Spark 1.2 within weeks, its first open release of the Muse era.
• Zuckerberg's same-day essay frames the move as anti-centralization strategy, aligning Meta with the open-weights letter camp.
• Every number so far is self-reported, and the flagship promise is unverified until the weights actually land.

Sources: Meta AI Research, Hugging Face model card, Mark Zuckerberg, "The Future Is For Everyone", Alexandr Wang on X, Techmeme (Wall Street Journal)