git clone https://github.com/newjordan/museB70.git
cd museB70

hf download Frosty40/Muse-Glimmer-30B-ArcB70-GGUF muse-glimmer-30B-kquant-17gb.gguf

gh release download v2026.08.12-b70 --repo newjordan/museB70 \
  --pattern 'muse-serve-3ce44d373-linux-b70.tar.zst'
sha256sum -c releases/ASSET_SHA256SUMS
tar --zstd -C releases -xf muse-serve-3ce44d373-linux-b70.tar.zst
rm -f muse-serve-3ce44d373-linux-b70.tar.zst

Serve

bash

MODEL=muse-glimmer-30B-kquant-17gb.gguf \
LLAMA_BIN=./releases/muse-serve-3ce44d373/bin/llama-server \
  ./serve-muse-arc.sh

bash

curl -s localhost:8095/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"model":"muse-glimmer-30b-q4","messages":[{"role":"user","content":"hi"}],"max_tokens":128}'

Muse thinks in reasoning_content. Echo both that and content on later turns, and give it enough max_tokens.

This is a serving configuration for the Intel Arc Pro B70 — not a new model.

Model and official kquant: meta-models/Muse-Glimmer-30B (GGUF), Apache-2.0, © Meta. Runtime: llama.cpp SYCL + Intel oneAPI. Serving config and B70 kernels: newjordan/museB70.