Muse Glimmer 30B

2026-08-10 · Publisher : Meta · Domain : Artificial Intelligence

Meta releases Muse Glimmer, a dense 30-billion-parameter open weights model under the Apache 2.0 license, distilled from Muse Spark and designed for local agents on a PC or Mac with a single GPU.

Meta releases Muse Glimmer, a dense 30-billion-parameter open weights model under the Apache 2.0 license, distilled from Muse Spark and designed for local agents on a PC or Mac with a single GPU. ### Context Since Muse Spark (April 2026), Meta had stopped releasing weights. By announcing Glimmer, Meta returns to open with a smaller model, distilled from its closed model, and hints that Muse Spark 1.2 will follow. The comparable open models of the period are Gemma 4 31B and Qwen3.6 27B. ### What the model brings - Always-on local agents: function calling, long tasks, error recovery, code and multimodal reasoning. - Logit distillation from Muse Spark during pre-training (Meta). - DFlash speculative decoding: up to 3.1 times faster on RTX 5090, 1.8 times on M5 Max, 1.5 times on M4 Max. - 4-bit quantization in K-Quant variants (17 GB) to fit on machines with 24 to 32 GB. - Dedicated perception encoder for image input. ### Architecture and specifications Dense transformer (not MoE), about 29.6B parameters including 1.8B for the vision encoder, 52 layers, attention alternating three local layers (2,048 window) and one global layer, GQA with 32 query heads / 2 KV heads, vocabulary of 202,048 tokens. Context: 131,072 tokens or more. Text and image input, text output, more than 100 languages. Knowledge cutoff: January 4, 2026. License: Apache 2.0 (free commercial use, no user threshold). Weights: Hugging Face meta-models/Muse-Glimmer-30B. ### Cost and access | Item | Value | |---|---| | Official API | Pricing not disclosed | | Memory | More than 55 GB at full precision; less than 20 GB of weights in 4-bit | | Measured throughput | 233.4 tokens/s (RTX 5090), 50.2 (M5 Max), 37.8 (M4 Max) with DFlash | | Ecosystem | llama.cpp, MLX, ExecuTorch, Ollama, Together AI announced | ### Benchmarks | Benchmark | Muse Glimmer 30B | Gemma 4 31B | Qwen3.6 27B | |---|---|---|---| | MCP Atlas | 75.5 | 54.2 | 62.5 | | DeepSearch QA | 74.6 | 61.7 | 71.1 | | SWE-Bench Verified | 76.0 | 66.6 | 77.2 | | SWE-Bench Pro | 51.2 | 36.9 | 50.2 | | AIME 2026 | 94.7 | 89.2 | 94.1 | | GPQA Diamond | 83.5 | 85.7 | 84.2 | Source: Hugging Face model card, scores published by the vendor; comparison with open models, not with a closed model. ### Versus closed models Meta does not compare Glimmer with closed models in the sources read: the published comparison covers two open models of similar size. It is ahead of Gemma 4 31B on most rows and on par with Qwen3.6 27B, with a clear advantage on MCP Atlas (75.5 versus 62.5) and a shortfall on SWE-Bench Verified (76.0 versus 77.2) and GPQA Diamond. The closed model from the same family, Muse Spark 1.2, is cited at 57 points on the Artificial Analysis index, with no published equivalent for Glimmer. The asset for the open versus closed choice is the Apache 2.0 license, more permissive than the Llama license, and local operation with no API cost or data transfer. ### Key takeaways Glimmer is the first open model of the Muse family: a dense 30B, strong in agentic tasks for its size, under Apache 2.0. It is still distilled from a closed model, and the promised release of Muse Spark 1.2 had not happened at the time of writing.

Tags : LLM, Open Source, Meta, Muse, Apache 2.0, Agents, Multimodal