Meta Muse Glimmer 30B Release

Meta has released Muse Glimmer, a 30B parameter multimodal model distilled from the Muse model and released under the Apache 2.0 license. Designed specifically for local agentic use cases, Muse Glimmer is optimized for privacy-aware applications including coding, document analysis, and personal assistants.

Technical Architecture

Muse Glimmer is a dense 30B parameter model composed of a 2B ViT-style Perception Encoder for vision and a 28B parameter text decoder.

Text Decoder Design

The text decoder utilizes several specialized architectural components to balance global context and efficiency:

  • Hybrid Attention: The model employs a repeating pattern of three sliding window layers (2,048 tokens) using rotary position embedding (RoPE), followed by a fourth layer using full attention and NoPE (no positional embedding). This pattern repeats 13 times across 52 total layers.
  • Gated Grouped-Query Attention (GQA): Each key-value head is shared by 16 query heads, reducing KV-cache memory requirements by 16x.
  • Q-K Normalization: RMS normalization is applied to every query and key head before attention computation to stabilize logits, followed by a scale factor applied to queries to act as an inverse temperature at the softmax level.

Perception Encoder and Multimodal Processing

The 2B ViT-like Perception Encoder handles both images and videos. It patchifies images into 2 frames x 3 channels x 14 x 14 shapes, applying interpolated absolute position embeddings from a learned table.

  • Vision Tower: Consists of 50 layers with GELU MLPs, utilizing a pattern of three window attention layers followed by one full attention layer with 2D RoPE.
  • Token Reduction: Pixel shuffle concatenates 2x2 groups of neighboring spatial tokens, reducing the image token count by 4x without losing channel information.
  • Video Processing: Videos are processed frame-by-frame at a target of 2 frames per second, capped at 96 frames. The system uses timestamped video placeholders (e.g., "Time: 0.0s <|video|>") to interleave text and video embeddings.

Performance Benchmarks

Muse Glimmer-30B demonstrates strong performance in agentic and multimodal tasks, often outperforming competitors like Gemma4-31B and Qwen3.6-27B in specific reasoning categories.

Category Benchmark Muse Glimmer-30B Gemma4-31B Qwen3.6-27B
General Agentic MCP Atlas 75.5 54.2 62.5
General Agentic GAIA2 43.3 36.4 40.0
Agentic Coding SWE-Bench Pro 51.2 36.9 50.2
Multimodal Charxiv Reasoning 78.8 77.7 78.4
General Reasoning AIME 2026 94.7 89.2 94.1
General Reasoning Beam 128K 65.1 58.2 63.0

Deployment and Inference Optimization

Speculative Decoding with DFlash

Muse Glimmer includes an optional speculative decoding drafter implemented on DFlash, a lightweight block-diffusion model. This drafter provides faster generation, particularly for structured content like code, by proposing up to 15 future tokens per step.

Ecosystem Support

The model features day-0 support for several major libraries and frameworks:

  • Transformers: Supports AutoModelForMultimodalLM and AutoProcessor across NVIDIA, AMD, and Intel GPUs.
  • llama.cpp: Supports calibrated quants and DFlash speculative decoding.
  • vLLM: Supported via the transformers backend.

Fine-Tuning Requirements

Fine-tuning can be performed using TRL (Transformer Reinforcement Learning) via SFT or Async GRPO. Hardware requirements vary by workload:

  • Inference/Eval (BF16): 1x 80GB H100.
  • LoRA SFT (BF16): 1x 80GB H100 (with microbatch 1 and checkpointing).
  • Full SFT (BF16): 8x 80GB H100 using FSDP/ZeRO-3.
  • LoRA GRPO: Requires 1x 80GB H100 (slow) or 8x H100 (4 for rollout, 4 for training).

Agentic Capabilities

Because of its multimodal and coding proficiency, Muse Glimmer can be configured as an autonomous agent capable of managing its own infrastructure. When connected to the Hugging Face MCP and OpenClaw, the model can perform the following tasks:

  • Self-Quantization: Search the Hub for GGUF weights, download them, or convert and quantize source weights using llama-quantize to run locally.
  • Self-Deployment: Deploy itself to Hugging Face Inference Endpoints using vLLM, verify health, and configure agent connections.
  • Self-Optimization: Benchmark its own serving stack on specific hardware (e.g., Nvidia H100) and iteratively test optimizations to maximize tokens/second throughput.

Sources

Related