Meta has released Muse Spark Open, an open-weight multimodal model, according to a Hugging Face launch post dated October 15, 2026. Hugging Face calls it the most capable US open-weights model to date.
What the model is
Muse Spark Open is a mixture-of-experts model with about 816 billion parameters in total and roughly 42 billion active for each token. Each layer holds 256 experts and the router picks 8 per token. It accepts text, images, audio and video, and produces text only. The context window reaches one million tokens, covering all input types. Meta trained it for agentic coding, meaning reading a codebase, writing and debugging code, and calling tools. It supports text-only and multimodal tool calling.
It is the full-size model of the Muse series. Muse Glimmer, a small dense model from the same series, came out in August and is meant for local use. Both share a tokenizer and vocabulary.
What it takes to run
All experts stay in memory, so the memory footprint is that of an 816B model even though speed resembles a 42B one. The BF16 checkpoint is about 1.5 TB. An NVFP4 checkpoint in 4-bit precision targets NVIDIA Blackwell GPUs, with what the post calls negligible accuracy loss.
The post lists several routes:
- Transformers: BF16 runs across four nodes of eight H200 GPUs; NVFP4 runs on a single B200 node.
- SGLang: supports text, image, audio and video; the reference NVFP4 setup uses four GB300 GPUs.
- vLLM: supports text and image only, with the NVFP4 checkpoint on Blackwell GPUs. The example config serves a 128K-token context.
- Inference Endpoints: a pre-validated NVFP4 recipe runs on eight RTX Pro 6000 GPUs at 131,072 tokens of context, or on eight H200 GPUs for the full 1,048,576. The endpoint exposes OpenAI-compatible Chat Completions and Responses APIs.
- Local Macs: on an M3 Ultra with 512 GB, a 2-bit llama.cpp quant used 304 GB and ran at 18 tokens per second. An MLX 3-bit quant took about 344 GB and generated about 21 tokens per second.
A full one-million-token context adds about 266 GB of KV cache in BF16, by the post's own calculation.
Reasoning and tuning
Reasoning cannot be switched off. Effort runs from low to xhigh, with xhigh as the default, and the post advises low for easier tasks to save tokens. For fine-tuning, Hugging Face describes a LoRA run on the attention layers with experts frozen, using TRL on eight nodes of eight H100 GPUs.
The excerpt does not cover licence terms or benchmark results, so check the model repositories for both before planning production use.
Top comments (0)