開放模型與推論系統
Muse Glimmer 30B Brings Multimodal Agents to a Single GPU With 17GB Quantized Weights and a DFlash Draft Model
Meta has released the Muse Glimmer 30B weights, vision encoder, and DFlash draft model. The 17GB version can run text, image, and speculative decoding workloads simultaneously within 24GB of VRAM. Meta measured an average of 233.4 tokens/s on an RTX 5090, but safety evaluations and deployment details show that the model still requires external safeguards and a newer runtime.
Meta has released Muse Glimmer, a roughly 29.6-billion-parameter model positioned not as a general-purpose chatbot but as an open-weight model capable of long-horizon reasoning, tool calling, error recovery, and screen and document understanding within local agent frameworks. Its architecture is a 52-layer dense Transformer with 32 query heads, 2 KV heads, and a repeating pattern of three local-attention layers followed by one global-attention layer. It is paired with a roughly 1.8-billion-parameter ViT-G/14 perception encoder and has a stated context length of 131,072 tokens.
The key deployment feature is that Meta directly provides two approximately 4-bit GGUF variants: a 16.8GB version targeting 24GB of VRAM and a 19.7GB dynamically quantized version targeting 32GB. The vision encoder requires another 1.4GB, while the five-layer DFlash draft model occupies 1.6GB. DFlash proposes 16 tokens at a time for parallel verification by the main model. At batch size 1 with greedy decoding, Meta measured an increase in average speed on an RTX 5090 from 74.9 to 233.4 tokens/s, while the M4 Max and M5 Max achieved 1.5× and 1.8× speedups, respectively. These are vendor measurements on specified hardware and prompt sets, not representative figures for multi-user serving throughput.
Compatibility also has explicit requirements: llama.cpp must be build b10353 or newer, following the August 10 merge that added the Muse Glimmer architecture. Older versions will fail to recognize the model entirely. Jinja must be enabled for the chat template, and the model’s reasoning channel cannot be disabled completely; users can only adjust the reasoning strength or set a token limit. If a server opens multiple slots, the context specified by `-c` is divided evenly among them, so long reasoning traces may exhaust the available context without producing an obvious error.
Official results list scores of 51.2 on SWE-Bench Pro and 75.5 on MCP Atlas, although the model still trails Qwen3.6-27B on benchmarks including OSWorld-Verified and TerminalBench 2.1. Its prompt-injection attack success rate on Siren AgentDojo also reaches 28.4%. Engineering teams should therefore treat it as an agent foundation model that can be deployed offline, rather than a finished product with comprehensive permissions management and security boundaries built in.