Muse Glimmer 30B

How it is built

Muse Glimmer counts about 29.6B parameters, and that figure includes the vision encoder.¹¹1Meta Superintelligence Labs, Muse Glimmer model card, Hugging Face, August 2026. https://huggingface.co/meta-models/Muse-Glimmer-30B The encoder alone is a ViT-G/14 of about 1.8B. Meta's card calls the model "distilled from Muse Spark" and built "for autonomous agentic tasks on consumer hardware".¹¹

It is a dense transformer. The language model has 52 layers with a hidden size of 6,656.¹¹ Three local layers come before every global one, which gives 39 local and 13 global (52 x 3/4 and 52 x 1/4).²²2Meta, config.json in the official repository: num_hidden_layers 52, layer_types, sliding_window 2048, layer_rope_theta 0 on every full_attention layer, max_position_embeddings 131072. https://huggingface.co/meta-models/Muse-Glimmer-30B/blob/main/config.json. Arithmetic: 52 x 3/4 = 39 sliding layers and 52 x 1/4 = 13 full layers, as layer_types lists them. The local layers attend over a sliding window of 2,048 tokens.¹¹

RoPE, with a base of 500,000, runs only in the local layers. In the config the global layers get no rotary position at all.²² Each layer has 32 query heads sharing 2 key-value heads, with gated attention.¹¹ The feed-forward block is SwiGLU, 19,968 wide.¹¹

What is new

The weights went public on Hugging Face on 10 August 2026, under Apache 2.0.³³3Hugging Face, launch post for Muse Glimmer, 10 August 2026, which says the model was "released today". This is the earliest official date; Meta's developer blog post is dated 12 August 2026. https://huggingface.co/blog/muse-glimmer Meta's developer blog calls it the most permissive licence Meta has used for an open model.⁴⁴4Meta for Developers, "Build with Muse Glimmer", 12 August 2026: "the most permissive license we've used for an open model"; "a default context window of 128K tokens". https://dev.meta.ai/resources/blog/build-with-muse-glimmer Llama 4 had come under the Llama 4 Community License.⁵⁵5Meta, Llama 4 Community License Agreement, effective 5 April 2025. https://github.com/meta-llama/llama-models/blob/main/models/llama4/LICENSE

For speculative decoding the model ships with a small DFlash drafter. It proposes 16 tokens in one forward pass, and the main model verifies them in parallel.¹¹

Meta measured the drafter with its 17 GB quantization at batch size 1. On an NVIDIA RTX 5090 under llama.cpp, generation went from 74.9 to 233.4 tokens per second.¹¹ An Apple M5 Max under ExecuTorch went from 26.6 to 50.2.¹¹

The comparison on the card is with Gemma 4 31B and Qwen3.6-27B, both thinking. At high reasoning Muse Glimmer scores 75.5 on MCP Atlas (Public), where the other two score 54.2 and 62.5. On SWE-Bench Verified it trails Qwen, 76.0 against 77.2, with Gemma at 66.6.¹¹

Running it

The repository has no access gate.⁶⁶6Hugging Face, repository metadata for meta-models/Muse-Glimmer-30B (gated: false), read 29 September 2026. https://huggingface.co/api/models/meta-models/Muse-Glimmer-30B At full precision the model targets 64 GB of VRAM. Meta compresses the weights to about 4 bits, which brings the language model under 20 GB.¹¹

K-Quant-Dynamic needs 32 GB and gives up 0.2% on average across 15 benchmarks. The 17 GB version fits in 24 GB, at a cost of 1.0%.¹¹

The config allows 131,072 positions, and Meta's blog gives 128K tokens as the default window.²²⁴⁴ Knowledge stops at 4 January 2026.¹¹ Reasoning strength is set in the system prompt, from low up to xhigh.¹¹

The Muse Spark models it was distilled from have no public weights.⁷⁷7Meta, Meta Model API documentation, which lists every Muse Spark version as an API model and Muse Glimmer as the self-hosted one, read 29 September 2026. https://dev.meta.ai/docs/models