1 · Connect your Hugging Face account
This Space hosts no weights — the NVFP4 model needs ~700 GB of VRAM, so it can't run on a Space GPU. Instead this page calls it through Inference Providers, and inference is billed to your Hugging Face account.
You'll need: (a) approved access to the gated model, and (b) a Hugging Face token with Inference Providers credits / a PRO plan.
Get a token at huggingface.co/settings/tokens · Request model access at meta-models/Muse-Spark-Open-816B-A42B-NVFP4 · Credits / billing at huggingface.co/settings/billing
Recommended sampling for this model: temperature 1.0, top_p 1.0 — those are used for every call.
2 · Chat
How this demo works
- Model:
meta-models/Muse-Spark-Open-816B-A42B-NVFP4— 816B total / 42B active, 62 layers, 256 experts (8 activated per token), 1M-token context, multimodal input. - Provider: Together, via the Hugging Face Inference Providers router (
router.huggingface.co/together/v1/chat/completions). Together serves the NVFP4 weights under the model idmeta-models/Muse-Spark-Open-816B-A42B— sending the-NVFP4name to Together gets "model not supported". - Why no hosted weights here: NVFP4 needs ~700 GB VRAM (BF16 needs ~2 TB). A Static Space is free for every account and needs no GPU — the model runs on the provider, not here.
- Your token never leaves your browser except to Hugging Face's router, in the Authorization header. No server in this Space stores it.
- Want the Gradio version?
gradio_app.pyin this repo is the same demo as a Gradio + ZeroGPU Space — use it once the account is 30+ days old, or on a PRO account.