Meta Muse Glimmer 30B

Meta's Open Weight — Muse Glimmer 30B

🎬 Sam Witteveen 📅 Aug 10, 2026 ⏱ 7:36
Meta Muse Glimmer 30B open weights Apache 2.0 distillation

🔓 Meta Returns to Open Weights

In what Sam Witteveen calls "a huge week for open models," Meta kicks it off with a release many didn't expect: Muse Glimmer — their return to releasing open-weight models. Zuckerberg himself announced Glimmer alongside confirmation that Meta will release weights for the larger Muse Spark 1.2 going forward. "While we're not seeing the return of Llama, it is awesome to have Meta back in the game contributing open weights." 0:00

The backstory: the original Muse model (April 2026) was a "very respectable model" from a team assembled late last year. Since then, Meta pulled off an under-reported coup — bringing over the head of reasoning from the Gemini team months before Google's high-profile exits began. That team has now delivered Spark 1.2, image models, video models, and now Glimmer. Released under Apache 2.0 — "even Yann LeCun, who left Meta in response to Alexander Wang's super-intelligence lab, is congratulating them." 0:50

📐 What Is Muse Glimmer 30B?

Glimmer is a 30B dense model (not mixture of experts), positioned as Meta's answer to Qwen 3.6 27B. The benchmarks show it beating Gemma 4 easily and trading wins with Qwen 3.6 across multiple categories — beating it on most but not all. Sam notes the timing: Qwen 3.8 27B is hours or days away — Meta may have rushed to ship before facing that comparison. 1:50

Built for agents: Meta states in the blog post that this model targets multi-step reasoning, tool use, long trajectories, and compatibility with coding harnesses, Open Claw, and Hermes Agent. "That is basically something that is going to have the ability for doing long trajectories and the ability to work with different harnesses." 4:09

🏋️ Training: Distillation + RL

The training approach is notable: Glimmer was trained via distillation on Muse Spark's output for pre-training — meaning it may not have been trained on raw internet data at all, unlike most models. "This is kind of interesting if this hasn't actually been trained on raw internet data or cleaned internet data like most of the models are using now." 3:15

Post-training follows a now-familiar pattern: a combination of on-policy distillation and reinforcement learning — similar to Thinking Machines' Inklings small model. "This is with a much smaller model and a model that people are going to be able to run locally." 3:51

💻 Running Locally — Quantized & Speculative Decoding

A major shift from the Llama era: Meta ships a 4-bit quantized version out of the box, sized to fit on 24 GB or 32 GB cards (3090/4090/5090, AMD 9700). "They've sized this model so not only does the model actually fit, but they're leaving quite a bit of headroom for a decent size KV cache." 4:48

Speculative decoding (D-Flash) is enabled, with demos showing it running on a MacBook Pro with 64 GB memory — performance that "just wouldn't have been the case with Llama models in the past." Token speeds shown for both Mac and 5090-class GPUs. Weights are already on Hugging Face. 5:18

📈 The Bigger Picture: Spark 1.2 Weights Coming

The most important signal: Zuckerberg confirmed that Muse Spark 1.2 weights are coming. On the Artificial Analysis Intelligence Index, Spark 1.2 is already showing as on par with Claude Opus 4.8. "Being able to run that locally is going to be a huge win for a lot of people." Combined with the Muse Code coding agent release, "all of these releases are really signaling that Meta is definitely back producing top-tier level models." 6:53

✅ Key Takeaways

  1. Meta is back in open weights with Apache 2.0 licensing. Muse Glimmer 30B marks their return — not Llama, but the Muse line, with Spark 1.2 weights promised.
  2. Dense 30B targeting Qwen 3.6 27B. Beats Gemma 4 handily, competitive with Qwen 3.6 on most benchmarks. Qwen 3.8 is imminent — Meta shipped fast.
  3. Trained via distillation, not raw internet data. Pre-training used Muse Spark's output; post-training combined on-policy distillation with reinforcement learning.
  4. Agent-first design. Purpose-built for multi-step reasoning, tool use, long trajectories, and harness compatibility (Open Claw, Hermes Agent).
  5. 4-bit quant out of the box, speculative decoding enabled. Fits on consumer GPUs with KV cache headroom. Runs on MacBook Pro 64 GB — a standard Llama-era models couldn't meet.
  6. Spark 1.2 weights coming — on par with Opus 4.8. If the Intelligence Index holds, running an Opus-class model locally will be a "huge win."

🔗 Resources & Links

📍 Timestamp Index

0:00 Meta returns to open weights — Muse Glimmer
0:50 Backstory: Gemini reasoning lead, Spark 1.2, Apache 2.0
1:50 30B dense model targeting Qwen 3.6 27B — benchmarks
3:15 Training: distillation from Spark + RL post-training
4:09 Agent-first design — tool use, long trajectories, harnesses
4:48 4-bit quant, speculative decoding, consumer GPU/MacBook
6:53 Spark 1.2 weights coming — on par with Opus 4.8
☰ View all