Meta has released Muse Glimmer-30B, a multimodal, open-source model designed for local, agentic use cases. The model is distilled from Muse down to 30 billion parameters and is published under the Apache 2.0 license, making it suitable for local deployment where privacy, cost reduction or hands‑on experimentation are priorities.
Intended uses
Muse Glimmer is positioned for privacy‑aware applications such as coding assistants, document analysis, personal assistants and agent frameworks modeled after Claw or Hermes. The release includes day‑0 support across several libraries to ease local use and integration.
Architecture overview
- Total size: 30B parameters — a 2B parameter ViT‑style Perception Encoder for vision and a 28B parameter text decoder.
- The text decoder uses a hybrid attention pattern: three sliding‑window attention layers (2,048 token windows with rotary position embeddings) followed by a full‑attention layer with NoPE (no positional embedding). The (SWA, SWA, SWA, Full) block is repeated 13 times for a total of 52 layers.
- Gated Grouped‑Query Attention shares each key‑value head across 16 query heads, reducing KV‑cache memory by 16× and reducing generation cost and latency.
- Q–K normalization with extra query scaling: queries and keys get RMS normalization and queries are scaled to control the logit scale (behaving like an inverse temperature at the softmax stage).
Perception Encoder (images and videos)
The Perception Encoder is a relatively large 2B ViT‑style encoder used for both images and videos. Inputs are patchified into a (2 frames × 3 channels × 14 × 14) shape, projected linearly, and receive interpolated absolute position embeddings before entering a 50‑layer vision tower with GELU MLPs. The attention pattern mirrors the language side with three windowed attention layers followed by a full attention layer and applies 2D RoPE to queries and keys. A pixel‑shuffle step concatenates neighboring spatial tokens (2×2 groups) to reduce token count by 4× while preserving channels. For video, the processor targets 2 frames per second and caps clips at 96 evenly sampled frames, inserting timestamped placeholders that are filled before the final projection.
Speculative decoding: DFlash
Muse Glimmer includes an optional DFlash speculative drafter that proposes blocks of future tokens to accelerate generation at some memory cost. Transformers and llama.cpp ship day‑0 support for the drafter; the authors highlight that the drafter is particularly useful for structured content generation such as coding.
Benchmarks — selected results
The release publishes benchmark scores as reported by the authors. Selected highlights include:
- General agentic: MCP Atlas — Muse Glimmer 75.5; DeepSearch QA — 74.6
- Agentic coding: SWE‑Bench Verified — 76.0; TerminalBench 2.1 — 51.7
- Multimodal: Charxiv Reasoning — 78.8; OmniDocBench v1.5 — 75.8
- Safety (CI Memories): Violation 26.4 (lower is better), Coverage 64.8
- General capabilities: IFBench 77.0; AIME 2026 94.7
The published table compares Muse Glimmer‑30B with models such as Gemma4‑31B and Qwen3.6‑27B; bolding in the original table marks the top result per benchmark.
Library and platform support
Day‑zero support is provided for several platforms and libraries, including:
- transformers: AutoModelForMultimodalLM and AutoProcessor classes (upgrade via pip install --upgrade transformers accelerate)
- llama.cpp: calibrated and community‑optimized GGUF quant files, server and CLI examples (llama serve meta-models/Muse-Glimmer-30B-GGUF)
- vLLM: support with transformers backend (examples for tensor‑parallel serving)
- Hugging Face Inference Endpoints: deployment and validation guidance
The model is available on the Hugging Face Hub under meta-models/Muse-Glimmer-30B.
Examples and demos
The release includes detailed examples for text‑only prompts, image+text prompting, multimodal tool calling (e.g., extracting a city from an image then calling a weather tool), open‑ended object detection output, and video question answering (with a 96‑frame cap). There are also demonstrations of speculative decoding speedups in transformers and llama.cpp.
Fine‑tuning and resource requirements
TRL‑based fine‑tuning examples are provided, from SFT to Async GRPO. The authors report experiments run in bf16 on 80 GB VRAM Hopper‑class GPUs with the following indicative resource needs:
- Inference/eval (bf16): 1 × 80 GB H100
- LoRA SFT (bf16): 1 × 80 GB H100 (microbatch 1 + checkpointing)
- Full SFT (bf16): 8 × 80 GB H100 with FSDP/ZeRO‑3
- LoRA GRPO and rollout configs: examples spanning 1 to 8 H100 setups
The release includes a MolmoWeb fine‑tuning example showing how to produce structured outputs and image fine‑tuning, and an OpenCode AsyncGRPO example for coding workloads.
Local deployment, quantization and agent workflows
The documentation shows how to find or create GGUF quantized weights (commonly Q4_K_M), run a llama‑server locally, and expose an OpenAI‑compatible /v1 API. It also provides AGENTS.md prompts used to let agents (OpenClaw, Hermes) search the Hub, select or create quantized weights, launch local servers, or deploy the model to Hugging Face Inference Endpoints and validate the endpoint.
Example agent capabilities include: discovering quantized weights on the Hub, switching to cached quantized weights, deploying the model to a protected Inference Endpoint, validating /health and /v1/chat/completions, and performing lightweight self‑optimization experiments (e.g., a protocol to maximize tokens/second on a single H100 while preserving correctness).
Wrapping up
Muse Glimmer‑30B is now available on the Hugging Face Hub. With its multimodal Perception Encoder, large text decoder, DFlash speculative drafter and day‑0 library support (transformers, llama.cpp, vLLM), it targets local, privacy‑sensitive, agentic applications such as coding assistants and personal agents. The model, demo notebooks and deployment examples can be found under meta-models/Muse-Glimmer-30B on Hugging Face.
