Meta has released Muse Glimmer: a 30B parameter local model that runs on consumer hardware
The new model is available on the Hugging Face Hub with immediate support in
transformers
,
llama.cpp
,
vLLM
, and Inference Endpoints.
Benchmarks
Test results
The following scores are published as reported. Bold text indicates the best result among the compared models. A downward arrow indicates that a lower score is better.
| Category | Benchmark | Muse Glimmer-30B | Gemma4-31B | Qwen3.6-27B |
|---|---|---|---|---|
| General Agentic | MCP Atlas | 75.5 | 54.2 | 62.5 |
| General Agentic | DeepSearch QA | 74.6 | 61.7 | 71.1 |
| General Agentic | τ³-Banking | 23.5 | 15.1 | 16.7 |
| General Agentic | WildClawBench | 47.6 | 37.6 | 43.2 |
| General Agentic | GDPval-AA | 953 | 811 | 1141 |
| General Agentic | GAIA2 | 43.3 | 36.4 | 40.0 |
| General Agentic | SkillsBench (With Skills) | 44.3 | 32.4 | 46.6 |
| General Agentic | OSWorld-Verified | 65.9 | 58.5 | 75.6 |
| 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 |
| Agentic Coding | SciCode | 43.6 | 43.4 | 39.8 |
| Multimodal | Charxiv Reasoning | 78.8 | 77.7 | 78.4 |
| Multimodal | ScreenSpot Pro | 75.4 | 75.9 | 76.1 |
| Multimodal | OmniDocBench v1.5 | 75.8 | 72.5 | 77.8 |
| Multimodal | MMMU Pro | 74 | 73 | 75 |
| Safety | CI Memories | Violation (↓): 26.4 Coverage: 64.8 | Violation (↓): 12.1 Coverage: 53.0 | Violation (↓): 53.4 Coverage: 66.9 |
| Safety | Siren AgentDojo | Attack Success Rate (↓): 28.4 Utility: 94.2 | Attack Success Rate (↓): 25.6 Utility: 90.8 | Attack Success Rate (↓): 40.3 Utility: 92.7 |
| General Capabilities and Reasoning | IFBench | 77.0 | 76.0 | 70.8 |
| General Capabilities and Reasoning | AIME 2026 | 94.7 | 89.2 | 94.1 |
| General Capabilities and Reasoning | GPQA Diamond | 83.5 | 85.7 | 84.2 |
| General Capabilities and Reasoning | Humanity’s Last Exam (Text + No Tools) | 22.0 | 23.6 | 23.1 |
| General Capabilities and Reasoning | AA-LCR | 80.0 | 68.3 | 73.3 |
| General Capabilities and Reasoning | Beam 128K | 65.1 | 58.2 | 63.0 |
Architecture
Muse Glimmer is a dense 30B parameter model consisting of:
- 2B ViT-style encoder for vision (Perception Encoder)
- 28B parameter text decoder
In addition to the main VLM, there is a speculative decoding drafter implemented on DFlash. Usage of this module is optional. It can provide much faster generation in exchange for some memory cost. The team found this drafter particularly well suited to structured content generation such as coding.
Text Decoder
The language model uses the following architecture components:
- Hybrid attention: Alternating between three sliding window layers (of 2,048 tokens) using rotary position embedding, followed by a fourth layer that uses full attention and NoPE (no positional embedding). The pattern is therefore (SWA, SWA, SWA, Full), repeated 13 times to a total of 52 layers. This allows the model to retain relative order and distance information with RoPE and preserve information globally with NoPE.
- Gated Grouped-Query Attention: Each key-value head is shared by 16 query heads, which reduces KV-cache memory by 16x and makes generation faster and cheaper.
- Q-K normalization with extra query scaling: Before computing attention, Muse Glimmer applies RMS normalization to every query and key head to keep attention logits stable. After this, queries are multiplied by a scale factor to set the target logit scale after normalization. The extra query scaling behaves like an inverse temperature at the softmax level.
Perception Encoder
Muse Glimmer uses one image encoder to handle both images and videos. Unlike the relatively small vision encoders used in other VLMs, this is a sizable 2B ViT-like model designed after the Perception Encoder architecture. Perception Encoder was previously introduced by Meta as a backbone for various downstream spatial and multimodal tasks.
The encoder patchifies images to a shape of 2 frames x 3 channels x 14 x 14, and passes them through a linear layer for projection. An interpolated absolute position embedding from a learned position table is then added to these embeddings. These are then sent to the vision tower which consist of 50 layers and GELU MLPs. Similar to the language model, the attention pattern consists of three window attention layers followed by one full attention layer. Inside the attention layers, 2D RoPE is applied to the queries and keys.
After transformer, pixel shuffle concatenates 2×2 groups of neighboring spatial tokens which reduces the number of image tokens 4x without discarding their channels. The merged features are then projected to the shared embedding space of the text decoder.
Videos go through the same encoder frame by frame, where each frame is converted into patches (of shape [batch, temporal groups, grid height, grid width, 2 frames, 3 channels, 14, 14]). The processor targets 2 frames per second and caps the clip at 96 frames sampled evenly across video. The processor creates timestamped video placeholders, interleaving text with frame e.g. “Time: 0.0s <|video|> x N” in which the final video embeddings are replaced before the final projection layer.
Transformers
Upgrade transformers to the latest version to be able to use Muse Glimmer.
pip install --upgrade transformers accelerateMuse Glimmer comes with day-0 support in transformers, both for the main model and the speculative decoding drafter. You can use
AutoModelForMultimodalLM
and
Source Read original →
