README.md

---
title: Muse Glimmer 30B
emoji: ✨
colorFrom: purple
colorTo: blue
sdk: gradio
sdk_version: 6.22.0
python_version: 3.10.13
app_file: app.py
startup_duration_timeout: 1h
fullWidth: true
pinned: false
license: apache-2.0
models:
  - meta-models/Muse-Glimmer-30B
  - meta-models/Muse-Glimmer-30B-assistant
---

# Muse Glimmer 30B · private inference

A private text-and-image chat for both:
- [`meta-models/Muse-Glimmer-30B`](https://huggingface.co/meta-models/Muse-Glimmer-30B), full BF16
- [`meta-models/Muse-Glimmer-30B-assistant`](https://huggingface.co/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:

```bash
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:

```bash
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:

```bash
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:

```bash
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`](https://huggingface.co/meta-models/Muse-Glimmer-30B/blob/f84ecc3a0ea984a4c04542a84269e3d065350a6e/USAGE_POLICY.md) or [`USAGE_POLICY.md`](https://huggingface.co/meta-models/Muse-Glimmer-30B-assistant/blob/2c86316d689027b91123638739743fef1d425233/USAGE_POLICY.md).
Model outputs may be inaccurate and should be independently verified for consequential use.