Meta Releases Muse Glimmer 30B, an On-Device Agentic Model with Built-In Perception Encoder

Meta Releases Muse Glimmer 30B for Local Agentic Workloads

Meta Superintelligence Lab has released Muse Glimmer-30B, a 29.6-billion-parameter causal language model with a dedicated perception encoder, distilled from a larger model called Muse Spark. According to Meta, the model is purpose-built for autonomous agentic tasks — multi-step reasoning, tool use, and multimodal understanding — running entirely on consumer hardware without cloud infrastructure. The model is released under an Apache 2.0 license with a training knowledge cutoff of January 4, 2026.

Architecture

Muse Glimmer is a dense causal transformer with 52 layers, a hidden dimension of 6656, and grouped-query attention using 32 query heads and 2 key/value heads (16:1 ratio). It uses a repeating local/local/local/global attention pattern with a 2048-token sliding window on local layers and RoPE positional encoding (θ = 500,000). The vocabulary spans 202,048 tokens. Context length is listed at 131,072+ tokens. A separate ~1.8B-parameter ViT-G/14 perception encoder (50 layers, 1536 width) handles image input, processing up to 4,096 visual tokens per image. Total parameters, including the vision encoder, are approximately 29.6B.

Local Deployment and Speed

Meta says the model is quantized to approximately 4-bit precision, compressing the language model to under 20 GB, allowing it to run alongside its KV cache, perception encoder, and a speculative decoding drafter within a 24–32 GB VRAM budget. Meta claims this quantization causes 0.2%–1.0% degradation on accuracy benchmarks compared to full precision, based on internal testing across 15 benchmarks.

The release includes a companion drafter model called DFlash, which predicts blocks of 16 tokens per forward pass for speculative decoding. Meta reports the following throughput on a 17GB quantized build: 233.4 tokens/sec on an Nvidia RTX 5090 (3.1x speedup over no-speculation baseline of 74.9 tok/s), 37.8 tokens/sec on Apple M4 Max (1.5x), and 50.2 tokens/sec on Apple M5 Max (1.8x).

Benchmark Results

Meta compares Muse Glimmer against Gemma4-31B (thinking mode) and Qwen3.6-27B (thinking mode). According to the company's published figures: MCP-Atlas 75.5 vs. 54.2 and 62.5; DeepSearch QA 74.6 vs. 61.7 and 71.1; SWE-Bench Verified 76.0 vs. 66.6 and 77.2; SWE-Bench Pro 51.2 vs. 36.9 and 50.2; AIME 2026 94.7 vs. 89.2 and 94.1; GPQA Diamond 83.5 vs. 85.7 and 84.2. On multimodal tasks, Muse Glimmer scores 78.8 on Charxiv Reasoning and 74 on MMMU Pro, broadly in line with the two comparison models. On agentic safety evaluations, Meta reports an AgentDojo attack success rate of 28.4% with 94.2% utility, versus 25.6%/90.8% for Gemma4-31B and 40.3%/92.7% for Qwen3.6-27B. These figures are self-reported by Meta and have not been independently verified.

The model supports four adjustable reasoning strengths (low, medium, high, xhigh) set via system prompt, and Meta states it is compatible with agentic orchestration frameworks including OpenClaw and Hermes Agent. Training data spans more than 100 languages.

What This Means

Muse Glimmer targets a specific niche: agentic AI that runs fully offline on a single consumer GPU or Apple Silicon Mac, rather than through a cloud API. The emphasis on quantization, speculative decoding, and VRAM budgets (24–32 GB) signals Meta is optimizing for developer workstations and high-end laptops rather than data centers. Benchmark comparisons against Qwen3.6-27B show mixed results — Muse Glimmer leads on some agentic tasks (MCP-Atlas, SWE-Bench Pro) but trails on others (SWE-Bench Verified, OSWorld-Verified, GPQA Diamond), suggesting no clear across-the-board advantage at this size class. Because it's Apache 2.0 licensed with no API listed, there is no per-token pricing; cost is effectively the hardware required to run it locally.