The weights are out; the superlative is Meta's
Meta's research post is dated October 15, with no clock. It says the weights for Muse Spark Open are out today, and calls it the most capable US open-weight model to date. The model is natively multimodal, with 816B total parameters, 42B active per token, and a 1M-token context, built for coding and long agentic work.
The developer post, the same day, calls it the open-weight Muse Spark, a sparse mixture of experts that activates roughly 42 billion parameters per token. It takes text, image, video, and audio, and holds a million tokens. The research post says 42B; the developer post says roughly 42 billion. Both stay as written. The superlative is not tied to an independent leaderboard in the post. The research post points to an evaluation-methodology note, and this page does not include the scores.
[1][2]
Two M5 Ultra machines, or four with 256GB each
The research post says this is the first model in its class to run on a desktop, using two Apple M5 Ultra Mac Studios and the latest MLX. The M5 Ultra's integrated GPU can use up to 512GB of unified memory per Mac Studio, so two machines can hold an entire model. Memory bandwidth is 1.2 TB/s. Meta says it tested agentic workflows on device, including shipping iOS changes to Ice Cubes, an open-source Mastodon client, and driving DaVinci Resolve end to end. That is their own description of a test, not a third-party rerun.
The developer post adds a path: the NVFP4 checkpoint can also sit on four Mac Studios with 256GB of unified memory each, 1TB in total. The weights do not move between the machines. Each layer is split, and each machine does half the computation.
[1][2]About 530GB, and the speedups are not a leaderboard rank
The developer post's table says the BF16 reference weights are about 1.6TB. NVFP4 is about a third of that, about 530GB. Only the experts are quantized. It serves on four Blackwell Ultra GPUs or eight H200s. The research post says the open-weight release ships auto-tunable Helion kernels in SGLang, with 1.03 to 3.95 times kernel-level speedups on NVIDIA and AMD GPUs, and a 3 percent end-to-end speedup on an AMD MI350. The developer post separately says SGLang's DFlash V2, on an AMD MI355X, gives 3.7 times the throughput for a single request and 2.9 times at 16 concurrent requests, with accuracy unchanged, and 6.4 drafted tokens accepted per step on average. Those are Meta's kernel and throughput figures, not a public leaderboard rank.
Hosted APIs named for launch day in the research post are Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. Weights are also on Hugging Face, Amazon SageMaker, Google Cloud Model Garden, and Microsoft Foundry. The license is the Muse Model License, with limited commercial terms. The post says many developers can use the model for free. Which commercial uses are excluded is in the license, not itemized in these two posts.
[1][2]The release bar was met; the full report is not out
The research post says the open weights got extra safety post-training, and that the model was evaluated under Meta's Superintelligence Scaling Framework for the most severe risks and met the thresholds that framework requires for release. Detailed results are in the model card. A full evaluation report is still to come. This post does not give the threshold scores, or name which risk categories were the binding ones.
[1]