Meta Muse Glimmer 30B: An Open-Weight Agentic Model Built to Run Locally

Meta Superintelligence Labs (MSL) announced Muse Glimmer on 10 August 2026 as the next model in its Muse line, open-sourcing the weights under Apache 2.0. Where the proprietary Muse Spark models target frontier-scale multimodal reasoning, Glimmer is the local-first member of the family: a 30-billion-parameter causal language model with a dedicated perception encoder, distilled from Muse Spark and built for autonomous agentic tasks on consumer hardware.

The model integrates multi-step reasoning, reliable tool use, multimodal understanding and failure recovery into a single checkpoint that runs without cloud infrastructure or network access. Meta positions it for always-on local agents, function calling, local coding and LLM-as-a-judge evaluation.

The core is a dense causal transformer with roughly 29.6 B parameters: 52 layers, a hidden dimension of 6656, SwiGLU feed-forward networks, and a repeating local/local/local/global attention pattern with a 2048-token sliding window. A frozen ViT-G/14 perception encoder adds about 1.8 B parameters and handles images, with a vocabulary of roughly 202,000 tokens and a context length of 131,072 (131 K+).

Input is text plus image; output is text. The knowledge cutoff is 4 January 2026, and the model was trained on data spanning more than 100 languages.

Muse Glimmer is optimised for local deployment. Quantising the weights to roughly 4-bit precision shrinks the language model to under 20 GB, leaving headroom for the KV cache, the perception encoder and a DFlash speculative-decoding drafter within a 24 GB or 32 GB envelope. Meta reports only 0.2% to 1.0% benchmark degradation across the quantised variants compared with full precision.

The DFlash drafter proposes blocks of 16 tokens in a single forward pass, which the main model verifies in parallel. On an RTX 5090, generation jumps from about 75 to 233 tokens per second (roughly 3.1x); on an M5 Max it rises from about 27 to 50 tokens per second (around 1.8x).

Meta evaluated Glimmer against Gemma4-31B and Qwen3.6-27B across agentic, coding, multimodal, safety and reasoning benchmarks. It leads the group on several agentic and reasoning tasks, including MCP Atlas (75.5), DeepSearch QA (74.6), Gaia2 (43.3), SWE-Bench Pro (51.2), AIME 2026 (94.7) and Beam128K (65.1). It does not win everywhere: Qwen3.6-27B posts higher scores on OSWorld-Verified, SWE-Bench Verified and TerminalBench, and Gemma4-31B leads on GPQA Diamond.

The practical takeaway for operators is that Glimmer is strongest where it is designed to be: end-to-end agentic task completion and long-context reasoning, rather than every single general-purpose benchmark.

Weights are available on Hugging Face at meta-models/Muse-Glimmer-30B, including full-precision BF16 weights, two 4-bit quantised variants, the DFlash drafter and the frozen vision encoder. Meta also published developer documentation under dev.meta.ai.

At the time of writing, the model is live on NVIDIA NIM (route meta/muse-glimmer-30b), which we verified with a successful request. OpenRouter, Together AI and Fireworks AI are named as coming partners, and integrations for Ollama, LM Studio, Unsloth, llama.cpp, MLX and ExecuTorch are expected in the days after release.