開放模型與端側推論
Meta Open-Sources Muse Glimmer 30B: 17GB Quantized Version with DFlash Hits 233 tokens/s on RTX 5090
Muse Glimmer brings multimodal input, tool calling, and failure retries to a 30B dense model deployable on consumer hardware. Apache 2.0 weights and official GGUF files are now available, although the performance and safety figures still come primarily from Meta’s own evaluations.
Meta released Muse Glimmer on August 10, a dense causal Transformer with approximately 29.6 billion parameters that supports text and image input, with its weights made available under the Apache 2.0 license. The model uses 52 layers, 32 query heads, and two KV heads with GQA. Its attention alternates between three local-attention layers and one global-attention layer, with a 2,048-token sliding window and a stated context length exceeding 131K. A ViT-G/14 perception encoder with approximately 1.8 billion parameters handles images, while output remains text-only.
The deployment focus is not full precision but two roughly 4-bit K-Quant variants. The smaller K-Quant-17GB can fit the language model into 24GB of VRAM; Meta says its average accuracy across 15 benchmarks drops by only 1%. The 32GB version reportedly loses just 0.2%. Meta also provides a five-layer DFlash draft model that proposes one 16-token block at a time for parallel verification by the main model. In Meta’s official batch-size-one, greedy-decoding tests, throughput on an RTX 5090 increased from 74.9 to 233.4 tokens/s, while throughput on an M5 Max rose from 26.6 to 50.2 tokens/s.
For agent evaluations, Meta reports scores of 75.5 on MCP Atlas and 51.2 on SWE-Bench Pro, respectively exceeding the Gemma4-31B and Qwen3.6-27B results listed by the company. However, Muse Glimmer still trails Qwen on OSWorld-Verified, TerminalBench 2.1, and some document-understanding tasks. Its safety table also shows a 28.4% attack success rate on Siren AgentDojo, underscoring that local execution is not inherently safe.
The weights, model card, and GGUF files are available for download, with launch instructions also provided for vLLM, SGLang, and Transformers. Optimized integrations with llama.cpp, MLX, ExecuTorch, Ollama, and others are marked as forthcoming. Engineering teams should next verify the combined memory requirements of the quantized files, vision encoder, and draft model, then rerun throughput and safety tests using their own agent harnesses, tool schemas, and prompt lengths.