Muse Glimmer 30B needs custom vLLM parsers

Muse Glimmer 30B is a dense vision-language model for agentic work on consumer hardware, combining a 52-layer text decoder with hidden size 6656 and a ~1.8B ViT-G/14 perception encoder. It has a 128K trained context, BF16 weights, Apache 2.0 licensing, a knowledge cutoff of January 4 2026, and training coverage across 100+ languages.

The model uses channel-scoped messages instead of JSON tool calls or <think> tags. vLLM requires both the muse_glimmer tool-call parser and reasoning parser, and the reasoning parser keeps skip_special_tokens=False so markers remain available for parsing. Supported checkpoints include BF16 at 59.58 GB, FP8 at 32.78 GB, NVFP4 at 25.42 GB, and a DFlash draft head at 5.11 GB.

Deployment guidance covers DGX Spark, GB300, GB10, and MI300X/MI325X/MI355X. On GB300 with TP=4, vLLM reports a 26.8M-token KV pool and ~204x max concurrency at the full 128K context. On MI355X, TP=4 reports a 27.68M-token KV pool and ~211.22x max concurrency.

Recommended sampling uses temperature = 1.0, top_p = 0.95, and top_k = 64, with reasoning strength set in the system prompt from low / medium / high / xhigh. Speculative decoding uses a DFlash block-diffusion draft head with num_speculative_tokens: 15, tied to the specific target model it was distilled against.