Meta Ships Muse Spark Open, an 816B-Parameter Open-Weight Frontier MoE
Meta released Muse Spark Open on October 15, an open-weight sparse MoE with 816 billion total parameters and 42 billion active per token, covering text, images, video, and audio with a 1M-token context. Weights are on Hugging Face with llama.cpp, MLX, SGLang, and vLLM support. Meta calls it the most capable US open-weight model, and early Apple Silicon quants already run it locally.
On this page
The release, and the claim
Meta released Muse Spark Open on 15 October, the open-weight version of its Muse Spark model: a sparse Mixture-of-Experts with 816 billion total parameters and 42 billion active per token, 256 experts per layer with 8 selected, and a 1M-token context window that covers text, images, video, and audio tokens together. Meta's team describes it as the most capable US open-weight model to date, coding-optimized and built for agentic work with multi-file refactors and debugging in training.
The weights are on Hugging Face in BF16, NVFP4, GGUF, and MLX quantizations, with day-one support in llama.cpp, SGLang, vLLM, and MLX, plus a fine-tuning recipe using TRL LoRA on 64 H100s. Coverage under the Muse Model License places it in the same family as the Muse Glimmer 30B dense model released in August, sharing its tokenizer and chat template while adding the new modalities.
Running it locally: the real numbers
The Apple Silicon numbers are the ones this audience asks about first, and the model card is specific: the team tested on an M3 Ultra with 512GB of unified memory, where a 2-bit llama.cpp quantization runs at about 18 tokens per second, and an oMLX 3-bit build of roughly 344GB reaches about 21 tokens per second for generation. Even the 3-bit quant needs more than 256GB of RAM, so this is Mac Studio class hardware, not a laptop model.
For data-center inference the requirements scale accordingly: BF16 needs four nodes of eight H200s, while the NVFP4 quantization fits on a single B200 node. The honest summary is that Spark Open is a frontier-adjacent model for prosumers and studios with unified-memory machines, not a consumer-hardware release; the model most people will run locally remains Glimmer 30B.
Where it sits in the open-weight race
The October field is suddenly crowded: Reflection's Beam (501B, Apache 2.0), Mistral Large 4 (1T, weights promised), and now Spark Open (816B, shipping) all landed within two weeks, each claiming a different edge: Beam inference efficiency, Mistral European sovereignty, Spark multimodal scale. Meta's announcement notably skips head-to-head benchmark tables against rivals like DeepSeek or Qwen, so the capability ranking is asserted rather than demonstrated.
What is demonstrated is the modality set and the tooling, and both are real: four input modalities in one checkpoint, a 1.8B perception encoder, and quantizations tuned for the machines people actually own. For local-AI users the decision is whether 42B active parameters of multimodal frontier fits a workload that Glimmer 30B cannot, at hardware prices that only studios pay. The weights are out; the benchmark tables are the part still to be proven.