Model releases · 11 August 2026
A 30-billion-parameter multimodal model built for agents that run on your own hardware, released with quantised GGUF weights, a speculative-decoding drafter and ExecuTorch builds for Apple silicon.
| Released by | Meta (Meta Superintelligence Lab) |
|---|---|
| Announced | 9–10 August 2026 |
| Parameters | ~29.6B total, including a ~1.8B vision encoder |
| Architecture | Dense causal transformer with a perception encoder (muse_glimmer) |
| Context length | 131,072 tokens |
| Modalities | Text and image in, text out |
| Licence | Apache 2.0, with a separate USAGE_POLICY.md in the repo |
| Weight format | safetensors, bfloat16 (59.6 GB across 2 shards) |
| Quantised builds | GGUF 16.8 GB and 19.7 GB; ExecuTorch .pte |
| Hosted at | meta-models on Hugging Face |
Muse Glimmer is a dense 30B model with a dedicated vision encoder, distilled from Meta's larger Muse Spark model and tuned specifically for agent work rather than open-ended chat. Meta's own description calls it "an open 30B model for always-on local agents."
The capabilities the model card puts forward are narrower and more specific than a general release: calling tools with correct schemas across long workflows, sustaining a plan over many steps, and — the one that stands out — diagnosing a failed tool call and retrying instead of stopping. It reports training and evaluation on DeepSearch QA, MCP-Atlas, 𝜏3-Bench and SWE-Bench, and support for more than 100 languages.
Mark Zuckerberg announced the release publicly and added that weights for Muse Spark 1.2, the foundation model Glimmer was distilled from, are also coming.
The configuration is unusual enough to be worth reading directly. The language model is 52 layers at hidden size 6656, with 32 query heads against 2 key-value heads — a 16:1 grouped-query ratio. Attention alternates three sliding-window layers (window 2048) with one full-attention layer, repeated thirteen times, and rotary position embeddings are applied only on the sliding layers.
The vision side is a 50-layer encoder at hidden size 1536, patch size 14, using windowed attention with periodic full-attention layers. The vocabulary is 202,048 tokens.
Four repositories, all under the meta-models organisation:
kquant-17gb at 16.8 GB and kquant-dynamic at 19.7 GB. Both are text-only on their own; image support needs the separate 1.4 GB mmproj projector in the same repo..pte files, in Metal and CUDA variants, across text-only / text+image and solo / drafter combinations.Meta reports the quality cost of the 4-bit compression itself as 0.2% degradation for the dynamic build and 1.0% for the 17 GB build.
The repository named -assistant is easy to mistake for a compact standalone version. It is not. Its configuration declares five layers and a target_layer_ids list pointing at layers 1, 13, 25, 37 and 49 of the 30B parent — it reads the larger model's hidden states and cannot run without it. There are no tokenizer files in the repo for the same reason.
What it does is block diffusion: it proposes an entire block of 16 tokens in one forward pass, and the 30B model then verifies those proposals in parallel, accepting the correct ones and correcting the rest. Meta's framing is that this produces identical output quality at higher speed.
Meta positions the release around a single GPU, and the file sizes match: the smaller GGUF build is sized to fit a 24 GB card, and the model card notes that its Apple silicon measurements were taken on M4 and M5 hardware using ExecuTorch, with the RTX numbers coming from llama.cpp.
| Runtime | Supported | Notes |
|---|---|---|
| llama.cpp | Partial | GGUF builds are published, but architecture support arrived as a separate pull request to ggml-org/llama.cpp after the weights. A build that predates it will reject the file with an unknown-architecture error. Both GGUF builds are text-only without the mmproj file. |
| ExecuTorch | Yes | Meta published compiled .pte files directly, including Metal variants for Apple silicon. |
| transformers | Partial | The config declares transformers 5.15.0.dev0, a development version rather than a release. |
| MLX | No | Not mentioned in any of the four repositories. |
| vLLM | Partial | Not officially listed. Community reports describe patching it manually to get the DFlash drafter working. |
USAGE_POLICY.md shipped alongside the Apache 2.0 licence has not been widely read; anyone building on the outputs should read it before assuming Apache terms apply end to end.Apache 2.0. The repository also contains a separate USAGE_POLICY.md file, which is worth reading before building on the model or its outputs.
The full bfloat16 weights are 59.6 GB across two safetensors shards. The quantised GGUF builds are 16.8 GB and 19.7 GB, plus 1.4 GB for the vision projector and 1.6 GB for the speculative-decoding drafter.
No. It is a five-layer drafter that reads hidden states from specific layers of the 30B model and cannot run on its own. It ships without tokenizer files for that reason.
Yes, through a 1.8B vision encoder. In GGUF form the main file is text-only; image input requires loading the separate mmproj projector alongside it.
Only with a build that includes the Muse Glimmer architecture support, which landed as a pull request after the weights were published. Older builds will refuse to load the file.
OnDevice LLM is a private AI assistant that runs entirely on your iPhone — no account, no cloud, and nothing you type leaves the device.
Published 11 August 2026