Intel / article

Muse Spark Open 816B A42B NVFP4 released

Source: huggingface.co
Author: Meta model releases
Date: Published Sep 15, 2026

Context

This is not a different model. It is Meta's Muse Spark Open, the same 816B-parameter model as the full-precision release, with its weights stored in NVFP4, a 4-bit number format made for NVIDIA's Blackwell GPUs. Storing each weight in fewer bits shrinks the memory the model needs: the model card lists at least 700GB of GPU memory for this version against 2TB for BF16 (model card).

That difference decides how many machines you need. Hugging Face's launch post says the BF16 checkpoint runs across four nodes of eight H200 GPUs, while this NVFP4 checkpoint runs on a single B200 node, and the post's vLLM and SGLang serving recipes both use this repository (Hugging Face). Hugging Face describes the accuracy loss as negligible; Meta's card is slightly more guarded, noting that may show minor quality differences in edge cases compared with full precision. Like the BF16 repository, downloads are gated behind the Muse Model License.

Terms in this piece · Glossary

mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.
quantization — Shrinking a model by storing its numbers less precisely — like rounding — so it runs faster and fits on smaller hardware, at a small quality cost.
inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.
multimodal — A model that works with more than text — reading images, audio, or video, and sometimes generating them too.

Read the source huggingface.co