Muse Glimmer 30B

NVFP4 · vllm · 256K context · RTX PRO 6000 Blackwell · 96 GB

Meta's Muse Glimmer 30B.

98 tok/sdecode

– tok/sprefill

256Kcontext window

Reported by MiaAI-LabAug 11, 2026

Run it

This is the launch as MiaAI-Lab publishes it. Their repository has the full setup: MiaAI-Lab/Muse-Glimmer-30B-DGX-Spark-RTX-5090-6000-PRO@1b71dc011010 › The server then answers on http://localhost:8888/v1.

1 Set up first

# Bare-metal (no container): needs python3.12, git, CUDA 13 at /usr/local/cuda; start.sh builds ./.venv with the vLLM fork, snapshot_downloads target + drafter into the default HF cache (MODEL_PATH/DRAFT_PATH are the returned snapshot dirs), runs `vllm serve` under nohup writing vllm.log/vllm.pid. 256K context (>131072) additionally REQUIRES two local, uncommitted patches to the fork (muse_glimmer.py and qwen3_dflash.py: size RoPE cos/sin cache to max_model_len) that are described in the README but NOT shipped in this repo; without them positions >131071 crash. Use --max-model-len 131072 otherwise.
# Full instructions: https://github.com/MiaAI-Lab/Muse-Glimmer-30B-DGX-Spark-RTX-5090-6000-PRO/tree/1b71dc011010a3e2c373eafd8a41f12c7c7913fc

2 Install

git clone https://github.com/MiaAI-Lab/Muse-Glimmer-30B-DGX-Spark-RTX-5090-6000-PRO && cd Muse-Glimmer-30B-DGX-Spark-RTX-5090-6000-PRO && git checkout 1b71dc011010a3e2c373eafd8a41f12c7c7913fc

3 Start the server

PATH='$VENV/bin:/usr/local/cuda/bin:$PATH' VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 '${VENV}/bin/vllm' serve '${MODEL_PATH}' --served-model-name muse-glimmer-30b --trust-remote-code --port 8888 --max-model-len 262144 --gpu-memory-utilization 0.46 --reasoning-parser muse_glimmer --tool-call-parser muse_glimmer --enable-auto-tool-choice --speculative-config '{"method": "dflash", "model": "${DRAFT_PATH}", "num_speculative_tokens": 16}' --kv-cache-dtype fp8_e4m3 --no-enable-flashinfer-autotune

What it passed

– loadLoads and serves within an hour

– chatAnswers a plain question and stops on its own

– reasoningThinks separately and gets 17 × 23 right

– toolsCalls a tool with the right arguments and uses the result

– contextRecalls a code buried in a prompt that fills 85% of the window

– speedDecodes at 15 tok/s or more

Published by MiaAI-Lab in MiaAI-Lab/Muse-Glimmer-30B-DGX-Spark-RTX-5090-6000-PRO@1b71dc011010, where it reports 97.9 tok/s. Our six checks have not run on it yet.

Details

Weights: inside the image
Image: none: a program on the host
Source: MiaAI-Lab/Muse-Glimmer-30B-DGX-Spark-RTX-5090-6000-PRO @ 1b7 ›
Launch: vllm-muse-glimmer-30b-nvfp4-256k-rtx-pro-6000-blackwell-96gb ›
Recipe file: registry/recipes/nvidia/rtx-pro-6000-blackwell-96gb/muse-glimmer-30b.vllm.256k.json ›
Model released: Aug 10, 2026
Tested: not yet by us; reported by miaai-lab