Meta shipped Muse Glimmer today: a 30-billion-parameter open-weight model, Apache 2.0, built to run agentic workflows locally on a single consumer GPU or a Mac. On its own, that's a solid, well-specified small-model release. What makes it worth a post is the sentence sitting next to it: Meta says Muse Spark 1.2 — the model this blog covered shipping closed five days ago — is going open too.
What Glimmer actually is
Muse Glimmer is a dense model — not MoE — built from a 27.9B text decoder, a 1.9B vision transformer, and a GELU multimodal projector, for roughly 30B total parameters. It takes interleaved text and images, so an agent built on it can read a screenshot or a chart alongside the conversation, not just text describing one.
The language model's attention is a repeating (sliding-window, sliding-window, sliding-window, full) pattern across 52 layers — three cheap local-attention layers for every one full-attention layer, with the full layer using NoPE (no explicit positional embedding) rather than RoPE. That's a deliberate cost trade: most of the network does short-range attention, and only every fourth layer pays for global context. Context length is 128K+.
The other real piece of engineering is DFlash, a small companion "drafter" network shipped alongside the main model. DFlash proposes an entire block of 16 tokens per forward pass; the main model verifies the block in parallel, keeping what's right and correcting what's wrong, rather than generating token-by-token. Meta reports 3.1× faster decoding on an RTX 5090, 1.8× on an M5 Max, 1.5× on an M4 Max — the gain shrinks on Apple silicon, consistent with a technique that pays off more where you have GPU parallelism to spend on verifying blocks than where you don't.
The training is a distillation, and the teacher is Muse Spark. Pre-training used logit distillation directly against Muse Spark's outputs, on a similar data mix to the teacher's own. Mid-training added longer-context, more agent-heavy data with richer reasoning traces. Post-training combined supervised fine-tuning with on-policy distillation and RL across general, reasoning, coding and agentic domains. That's the same instinct behind Muse Spark 1.1 generating and grading its successor's training data, run across a size class instead of a model generation: the flagship Meta just kept closed is the thing that taught the model Meta just opened. Meta also says Glimmer was evaluated under its own Advanced AI Scaling Framework before this release — a named internal gate for open-weight releases, the same shape of thing OpenAI's Preparedness Framework is for capability thresholds, applied here to justify shipping open rather than to justify shipping at all.
At full precision the 30B model needs over 55GB of memory — no consumer GPU has that. Meta quantizes the weights to roughly 4-bit, shrinking the language model to under 20GB, which leaves enough headroom for the KV cache, the perception encoder and the quantized DFlash drafter to all run together inside a 24–32GB card. Meta states this compression costs minimal to no degradation on agentic tasks; that claim comes from the same vendor as the benchmarks below, so read it with the same discount.
Performance: a plurality, not a sweep
Meta's own model card publishes 24 benchmark rows across general agentic tasks, agentic coding, multimodal understanding, safety, and general reasoning — evaluated at Glimmer's High Reasoning setting against Gemma4-31B and Qwen3.6-27B in Thinking Mode, so all three are being read at their strongest configuration. The picture that emerges is more interesting than "Glimmer wins":
| Category | Benchmark | Muse Glimmer | Gemma4-31B | Qwen3.6-27B |
|---|---|---|---|---|
| General agentic | MCP Atlas | 75.5 | 54.2 | 62.5 |
| General agentic | GAIA2 | 43.3 | 36.4 | 40.0 |
| General agentic | OSWorld-Verified | 65.9 | 58.5 | 75.6 |
| General agentic | SkillsBench (with skills) | 44.3 | 32.4 | 46.6 |
| Agentic coding | SWE-Bench Pro | 51.2 | 36.9 | 50.2 |
| Agentic coding | SWE-Bench Verified | 76.0 | 66.6 | 77.2 |
| Agentic coding | TerminalBench 2.1 | 51.7 | 43.4 | 60.7 |
| Multimodal | OmniDocBench v1.5 | 75.8 | 72.5 | 77.8 |
| Safety | Siren AgentDojo (attack success, lower better) | 28.4 | 25.6 | 40.3 |
| Reasoning | GPQA Diamond | 83.5 | 85.7 | 84.2 |
| Reasoning | AA-LCR | 80.0 | 68.3 | 73.3 |
Tallied across all 24 published rows, Glimmer leads on 12, Qwen3.6-27B leads on 8, Gemma4-31B leads on 4. That's a real edge, not a rounding error — but it isn't the sweep the two headline benchmarks in Meta's own blog post (MCP Atlas, SWE-Bench Pro) would suggest on their own, and it clusters in a legible way. Glimmer's wins concentrate in end-to-end agentic task completion and general reasoning — exactly what a model distilled from a flagship on agent-heavy, reasoning-rich data should be good at. Qwen3.6-27B's wins concentrate in sustained, verified execution — OSWorld-Verified, SWE-Bench Verified specifically rather than the noisier Pro variant, TerminalBench, document parsing — the kind of long, must-not-drift-off-track tasks where an open model with a longer public track record might simply have had more scrutiny and iteration. And Gemma4-31B, despite trailing almost everywhere else, wins the attack-resistance benchmark and the hardest pure-knowledge benchmark (GPQA Diamond), suggesting a different point on the safety/capability trade-off rather than a strictly weaker model.
Vendor-selected panel, vendor-run numbers — same caveat this blog applies to every self-graded chart. But a 24-row table with losses in it is a more trustworthy document than the two rows that made it into the announcement post, and it's worth reading the whole thing rather than the highlight reel.
The sentence that matters more than the model
Five days ago this blog covered Muse Code and Muse Spark 1.2 — Alexandr Wang's first model release leading Meta's coding line, shipped with no open-weights option at all, completing what the arcade's own dataset had been tracking as a reversal from Meta's Llama-era identity stretching back to OPT-175B in 2022. That post, after a direct instruction from the person this blog is written for, does not editorialize about what that reversal might mean — the editorial framing was removed because it wasn't part of what Meta had actually said that day.
Today Meta said something. Alongside Glimmer's launch, Zuckerberg published a roughly 14-page essay, "The Future is for Everyone," confirming that Muse Spark 1.2 will get an open-weight release. Alexandr Wang is quoted the same way. No license, no date, no parameter count for the open version — just the commitment. It's worth noting Zuckerberg had already been asked about it directly: on X, the day Muse Spark 1.2 shipped closed, he replied to a question about open weights with "I'll have more to share on that soon." Five days is a plausible reading of "soon."
Treat this the way this blog treats every unshipped announcement: a stated intention, not a release. Meta's own recent history includes models that stayed open (Muse Spark, LLaMA), models that shipped closed and stayed closed (Muse Image, Muse Video, Spark 1.1), and now a model that shipped closed and — on the company's own word — won't stay that way. The right entry for "Muse Spark 1.2: open" is the day weights land on Hugging Face with a license attached, not the day someone says it's coming.
The other half of the essay
The open-weights announcement isn't happening in a vacuum. Zuckerberg's essay frames it explicitly as competitive positioning against Chinese open-weight labs — Alibaba, DeepSeek, Moonshot — arguing that "foreign labs currently hold several advantages" because American labs face training-data restrictions their competitors don't, and that "US policy must reduce this additional friction if we want American open source models to lead over time." That's the same fight this blog covered in July, when a floated Washington ban on Chinese open-weight models drew a fast, unusual industry coalition against it. Zuckerberg's essay is Meta arguing the reverse side of that same policy fight: not "block their open models," but "loosen the rules on ours so we can out-compete on openness instead."
The essay also announces a $1 billion community fund aimed at communities near Meta's data-center buildout — a separate, non-model commitment bundled into the same release, which is its own signal about how much political weight this announcement is meant to carry.
What to expect next
- Watch for the actual Spark 1.2 weights, not the essay. The gap between "we will open source this" and a downloadable checkpoint with a real license has already been five days once this year; there's no reason to assume it closes quickly.
- Watch which rung of openness it lands on. Nothing about "open" is one thing — Llama's threshold-gated license, Apache 2.0 like Glimmer's, or something in between are three very different commitments wearing the same word.
- Glimmer is the more durable story regardless of what happens to Spark 1.2. A real Apache 2.0 local-agent model with competitive numbers against Gemma4 and Qwen3.6 changes what runs on a consumer GPU today. A promise about a future release changes nothing until it ships.
References: Meta AI Research — Introducing Muse Glimmer · Muse-Glimmer-30B model card — Hugging Face · CNBC — Meta launches Muse Glimmer open-weight AI model · Phoronix — Meta Publishes Muse Glimmer as 30B Open Agentic Model · related coverage: Meta's Muse Code Bets on Agents That Don't Forget · The Week Washington Flirted With Banning Open Weights · How Open Is 'Open'? · The Heavyweight Shift