ssdataanalysis/muse-glimmer-30b

Muse Glimmer 30B · private inference

A private text-and-image chat for both:

• meta-models/Muse-Glimmer-30B, full BF16
• meta-models/Muse-Glimmer-30B-assistant, assistant checkpoint

meta-models/Muse-Glimmer-30B-assistant is a companion DFlash drafter and does not expose standalone .generate(...) in this release of Transformers. In this Space, assistant selection runs the full 30B checkpoint (same prompt/history behavior) while keeping the model selector and UI behavior intact.

They are served at pinned revisions: f84ecc3a0ea984a4c04542a84269e3d065350a6e (full BF16, 29,776,626,688 parameters) and 2c86316d689027b91123638739743fef1d425233 (assistant BF16, 2,555,985,152 parameters).

Runtime

• Full, unquantized BF16 checkpoint (29,776,626,688 parameters)
• ZeroGPU allocation through the @spaces.GPU(size="xlarge") request. The full 30B checkpoint reserves up to 120s for its cold start on first load by default. Assistant selection follows the same warm-path budget rules:
  • if a full model is already loaded in RAM, assistant generation reuses that active model and uses the normal per-turn estimate,
  • if assistant must warm an alternative path, it also uses the 120s cold-start cap.
• Pinned read-only model storage is mounted from Hugging Face model volumes into
  • /models/muse-glimmer
  • /models/muse-glimmer-assistant
• Runtime defaults can be overridden with MUSE_MODEL_PATH, MUSE_ASSISTANT_MODEL_PATH, and MUSE_DEFAULT_MODEL_ID.
• transformers==5.15.0, the first stable release with native Muse Glimmer support
• Sequential safetensors pread loading, avoiding whole-shard host-RAM copies from the volume
• One inference request at a time
• Optional bucket-backed checkpoints are supported by setting MUSE_MODEL_PATH and MUSE_ASSISTANT_MODEL_PATH to mounted bucket locations (for example /mnt/muse-glimmer when mounted via hf://buckets/<namespace>/<bucket>:/mnt).
• The private Space API path is locked down (api_visibility: private) so all chat requests flow through the same Gradio queue and preflight validation.
• Model loading is deferred until the first request so the Space can start even when a mount is temporarily unavailable. To keep both checkpoints selectable, mount both model volumes in Space settings: /models/muse-glimmer and /models/muse-glimmer-assistant.

Optional Hugging Face bucket storage

If you need storage that persists beyond the Space runtime, mount an HF bucket:

hf spaces volumes set ssdataanalysis/muse-glimmer-30b \\
  -v hf://buckets/ssdataanalysis/muse-glimmer-30b-store:/mnt/muse-glimmer-store \\
  -v hf://models/meta-models/Muse-Glimmer-30B:/models/muse-glimmer:ro \\
  -v hf://models/meta-models/Muse-Glimmer-30B-assistant:/models/muse-glimmer-assistant:ro

Then point the app to bucket copies without changing code:

export MUSE_MODEL_PATH=/mnt/muse-glimmer-store/Muse-Glimmer-30B
export MUSE_ASSISTANT_MODEL_PATH=/mnt/muse-glimmer-store/Muse-Glimmer-30B-assistant

To inspect configured mounts:

hf spaces volumes list ssdataanalysis/muse-glimmer-30b

The app loads the selected checkpoint on demand; a selected model switch swaps checkpoints as needed. Loaded checkpoints are placed on cuda, as required by ZeroGPU. The app does not download full checkpoints into the Space's ephemeral disk.

ZeroGPU status check:

• hf spaces info ssdataanalysis/muse-glimmer-30b --json should show:
  • requested_hardware: zero-a10g
  • hardware: zero-a10g
• This app requests @spaces.GPU(size="xlarge") and should run on ZeroGPU's 96GB path for BF16.

Runtime access and quota checks

• The Space is private. If you open https://huggingface.co/spaces/ssdataanalysis/muse-glimmer-30b or https://ssdataanalysis-muse-glimmer-30b.hf.space without being authenticated as the owner, you will see an authentication/error page.
• API calls must include an authenticated Hugging Face token with access to this private Space.
• You can verify the endpoint is wired with:

HF_TOKEN=<your-token>
curl -H "Authorization: Bearer $HF_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"data":["hi",null,"meta-models/Muse-Glimmer-30B",[],[], "", "high", false, 32, 1.0, 0.95, 64, 1.0, 42, false, true]}' \
  https://ssdataanalysis-muse-glimmer-30b.hf.space/gradio_api/call/chat

• If you receive a ZeroGPU quota response such as You have exceeded your Pro ZeroGPU quota ..., the container/startup is healthy and the failure is account-level quota, not a model loading/runtime regression. Add credits here if you need immediate inference: https://huggingface.co/settings/billing?add-credits=true

Defaults and controls

The checkpoint's generation_config.json uses greedy decoding (do_sample: false), so the UI opens in Native greedy mode. The separate Meta recommended sampling preset enables the model-card values:

• temperature: 1.0
• top-p: 0.95
• top-k: 64

Reasoning strength defaults to high, matching the supplied chat template. It can be changed to low, medium, high, or xhigh. The app also exposes the response token budget, sampling toggle, repetition penalty, seed, optional system instruction, reasoning visibility, and optional image input. The 512-token response budget is an app safety default; the model repository does not define a practical per-response budget.

The app uses the model's native response parser to keep reasoning and final-answer regions separate. No tools are registered or executed.

Privacy and safety

Access is enforced by the private Space setting. This application does not write prompts, responses, or uploaded images to persistent storage and does not print them to application logs. Inference still runs on Hugging Face-hosted infrastructure; it is not local-only. Image metadata is discarded before inference.

Use is subject to the model's USAGE_POLICY.md or USAGE_POLICY.md. Model outputs may be inaccurate and should be independently verified for consequential use.