Hardware platforms

Run Muse Glimmer on specific partner hardware. Each partner verified its deployment on its own machines at the software versions listed; we don't re-run these verifications. Pick a platform to jump to its deploy steps.

Vendor	Platform	Class	Accelerator	Memory	Precisions verified
Intel	Arc Pro GPU (server)	Server	Arc Pro B70 × 2	32 GB per accelerator	bf16
Intel	Xeon 6 6980P	Server	CPU only	More than 512 GB DDR5 suggested	bf16
NVIDIA	B300	Server	B300 × 1 (8-GPU node)	288 GB HBM3e per GPU	bf16

SGLang publishes its own matrix of verified GPUs and Apple silicon; see Choose a checkpoint on Run inference.

Intel

Intel verified Muse Glimmer with vLLM on Arc Pro GPUs and Xeon processors. Intel's day-0 guide, Day-0 Muse Glimmer Support on Intel Platforms with vLLM and Hugging Face(opens in new tab), adds the tool-calling flags and a Hugging Face Transformers alternative, including DFlash speculative decoding on Xeon.

Arc Pro GPU (server)

Accelerator	Arc Pro B70 × 2
Memory	32 GB GDDR6 per card
Memory bandwidth	608 GB/s
Host OS tested	Ubuntu 24.04
Driver and runtime	Level Zero 1.28.2 (check with `xpu-smi`)
Model server	vLLM `main` branch
Disk	80 GB free for the bf16 weights
Verified	Intel, 2026-08-09

Build and enter the vLLM XPU container:

bash
git clone https://github.com/vllm-project/vllm.git
cd vllm
docker build -f docker/Dockerfile.xpu -t vllm-xpu-env --shm-size=4g .
docker run -it --rm --network=host --ipc=host --privileged \
                 --device /dev/dri:/dev/dri \
                 -v /dev/dri/by-path:/dev/dri/by-path \
                 --entrypoint bash \
                 vllm-xpu-env

Download the weights and serve across both cards:

bash
hf download meta-models/Muse-Glimmer-30B --local-dir ./muse-glimmer-bf16
ZE_AFFINITY_MASK=0,1  \
python -m vllm.entrypoints.openai.api_server \
  --model ./muse-glimmer-bf16 \
  --served-model-name muse-glimmer \
  --reasoning-parser muse_glimmer \
  --enforce-eager \
  --tensor-parallel-size 2

Tensor parallelism across more accelerators, such as --tensor-parallel-size 4, also works. Intel reports validating this path on Arc Pro B70 and Arc Pro B60.

Precision	Status	Artifact
bf16	Verified	`meta-models/Muse-Glimmer-30B`

Xeon 6 6980P

Accelerator	None, CPU only
Memory	Depends on the DDR5 installed; more than 512 GB suggested
Memory bandwidth	1228.8 GB/s
Host OS tested	Ubuntu 24.04
Model server	vLLM `main` branch
Disk	80 GB free for the bf16 weights
Verified	Intel, 2026-08-09

Build and enter the vLLM CPU container:

bash
git clone https://github.com/vllm-project/vllm.git
cd vllm
docker build -f docker/Dockerfile.cpu \
        --tag vllm-cpu-env \
        --target vllm-openai .
docker run -it --rm \
            --privileged \
            --shm-size=4g \
            --entrypoint bash \
            --name vllm-cpu-env \
            vllm-cpu-env

Download the weights and serve:

bash
hf download meta-models/Muse-Glimmer-30B --local-dir ./muse-glimmer-bf16
vllm serve ./muse-glimmer-bf16 \
    --served-model-name muse-glimmer \
    -tp "<TP_SIZE>" \
    --dtype=bfloat16 \
    --trust-remote-code --enable-auto-tool-choice \
    --tool-call-parser muse_glimmer \
    --reasoning-parser muse_glimmer \
    --generation-config auto

Confirm the deploy with the vLLM tool-calling check.

Precision	Status	Artifact
bf16	Verified	`meta-models/Muse-Glimmer-30B`

For Arc Pro and Xeon issues, use the vLLM issue tracker(opens in new tab) and the vLLM installation docs(opens in new tab).

NVIDIA

NVIDIA data-center GPUs run Muse Glimmer with vLLM on the standard CUDA stack, with no plugins. The serving path is the same as the vLLM section. Once vLLM is serving, any agent harness can drive the OpenAI-compatible endpoint. NVIDIA's NemoClaw(opens in new tab) is one: it runs the agent loop in a sandbox with filesystem, network, and credential policy in front of the tools.

B300

Accelerator	B300 SXM6 AC × 8 (compute capability 10.3)
Memory	288 GB HBM3e per GPU (275,040 MiB reported)
Memory bandwidth	About 8 TB/s per GPU
Host OS tested	Ubuntu 24.04.3 LTS
Driver and runtime	Driver 580.167.08 or later (580 or later for Blackwell Ultra sm_103), CUDA 13.0
Python	3.10 or later (tested on 3.12.3)
Model server	vLLM `0.1.dev19055+g6b0c510ec`, torch 2.13.0+cu130, transformers 5.14.1
Disk	80 GB free for the bf16 weights
Verified	NVIDIA, 2026-08-10

bash
hf download meta-models/Muse-Glimmer-30B --local-dir ./muse-glimmer-bf16

bash
docker run --rm --gpus all --ipc=host \
  -p 8000:8000 \
  -v "$PWD/muse-glimmer-bf16:/model" \
  vllm/vllm-openai:muse-glimmer \
  /model \
  --served-model-name muse-glimmer \
  --tensor-parallel-size 1 \
  --enable-auto-tool-choice \
  --tool-call-parser muse_glimmer \
  --reasoning-parser muse_glimmer \
  --generation-config auto

The bf16 weights fit on a single B300, so --tensor-parallel-size 1 is correct on an 8-GPU node too. NVIDIA took its measurements on the vLLM build in the spec table rather than this container; the flags are the same. Confirm the deploy with the vLLM tool-calling check.

Precision	Status	Artifact	Notes
bf16	Verified	`meta-models/Muse-Glimmer-30B`	Fits on a single B300.

NVIDIA measured these single-GPU numbers on 2026-08-09. The node has eight B300s, but the sweep used one, without tensor parallelism:

Concurrency	TTFT p50	Output tok/s per request	Output tok/s total
1	215.6 ms	85.76	76.83
2	307.3 ms	84.44	148.81
4	395.3 ms	81.99	280.96
8	618.7 ms	77.37	503.27
16	979.7 ms	67.26	830.71
32	1747.9 ms	53.29	1,224.67
64	2760.0 ms	36.53	1,632.21
128	3432.2 ms	19.69	1,912.99
256	2790.0 ms	9.88	2,093.52

The run used vLLM 0.1.dev19055+g6b0c510ec, torch 2.13.0+cu130, driver 580.167.08, and CUDA 13.0, with 2048 input tokens, 256 output tokens, and a 32768-token prefix cache. Reproduce it with:

bash
aiperf profile --model '/model' --tokenizer '/model' --url 'http://localhost:8000' \
  --endpoint-type 'chat' --streaming \
  --num-prefix-prompts 1 --prefix-prompt-length 32768 \
  --num-dataset-entries 10240 --dataset-sampling-strategy 'shuffle' --random-seed 7 \
  --isl 2048 --isl-stddev 0 --osl 256 --osl-stddev 0 \
  --concurrency '1,2,4,8,16,32,64,128,256' \
  --request-count '10,20,40,80,160,320,640,1280,2560' \
  --sweep-type 'zip' --warmup-request-count 3 --num-profile-runs 1 \
  --use-server-token-count --use-legacy-max-tokens \
  --extra-inputs 'max_tokens:256' --extra-inputs 'min_tokens:256' \
  --extra-inputs 'ignore_eos:true' --extra-inputs 'temperature:0'

Troubleshoot a deployment

These fixes apply on every platform above.

Symptom	Cause	Fix
Tool calls returned as plain text	Parser not enabled	Pass both `--enable-auto-tool-choice` and `--tool-call-parser muse_glimmer`. The chat template ships with the checkpoint, so no `--chat-template` flag is needed.
Generation never stops	The runtime overrode the stop tokens	Confirm `<\|eom\|>` isn't in `eos_token_id`, and serve with `--generation-config auto`.
Out of memory at large batch sizes	KV cache exhausted	Lower `--max-num-seqs` or `--max-model-len`.

Next steps

For every runtime and its full flag set, see Run inference. To shrink the model for smaller hardware, see quantization.