On October 15, 2026, Meta released Muse Spark Open, which its announcement calls "the most capable US open-weight model to date". The model card describes "an 816B total/42B active parameter open weights model with a context window of 1M tokens", natively multimodal, with text, image, video and audio input and text output. Weights ship in BF16 and NVFP4 on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models, and hosted APIs are available from Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. It follows the closed Muse Wonder launch on October 12, and Meta keeps its strongest models closed.
The desktop pitch
Meta's headline demo is local: the model "is the first model in its class to run on your desktop using two Apple M5 Ultra Mac Studios with the latest MLX framework", using up to 512 GB of unified memory per machine. That is the Apple path. For GPUs, the model card says it "requires a GPU with at least 700GB of VRAM for NVFP4 and 2TB for BF16". Meta's developer post adds custom Helion kernels and reports speculative decoding giving "3.7× throughput for a single request on AMD MI355X".
Safety post-training for an open model
The card says "This open weights version has additional safety post-training to make it more robust against misuse as a standalone model without system guardrails." Meta reports that with no system-level safeguards, it "refused 99.2% of jailbreaks on our internal benchmark and 94.8%" across three public cyber-misuse benchmarks, refused "over 98% of harmful requests on SecureBio's BioTIER refusal benchmark", and showed "a 98.1% honesty rate on the MASK benchmark and a 0% misbehavior rate on the Agentic Misalignment benchmark." These are Meta's own numbers; a full evaluation report is promised "soon".
The license is the story for builders
Meta says the Muse Model License has "limited commercial terms, which allow many developers to use the model for free." Two terms matter. Products must "prominently display 'Built with Muse Spark Open'". And if a licensee offering the model as an API service has "aggregate gross revenue" above "twenty (20) million US dollars" in any 12-month period, it "must notify Meta and enter into a separate written agreement" first. Internal use, research, and powering one's own product are excluded.
That clause drew the sharpest response in a Hacker News thread that passed 670 points. User popinman322 wrote: "The usual effect of separate licensing is decreased price competition from public providers". On capability, commenters pointed to an Artificial Analysis Intelligence Index score of 39, which user GodelNumbering said is "about the same as Deepseek v4.1 flash and Mimo 2.5 Flash" and behind GLM 5.3 Flash and Qwen3.8-Flash-Next.
What remains uncertain. Independent evaluations are only starting, and the score comparisons above come from commenters, not from a published study. Meta has not said what terms large providers will be offered.
Analysis: "Most capable US open-weight model" is a narrower claim than it sounds: the US qualifier does the work, because Chinese open models of a fraction of the size score as well or better. The practical value is elsewhere. A frontier-lab model with long context, wide engine support and published safety post-training is useful to companies that need a US-origin model for procurement or compliance reasons. Teams planning to resell inference should read Section 3 of the license before building on it.
Meta released the weights of Muse Spark Open on October 15, an 816B-parameter (42B active) multimodal model with a 1M-token context that runs on two M5 Ultra Mac Studios; its license requires API providers above $20 million in revenue to sign a separate deal, and early readers placed its Artificial Analysis score of 39 behind smaller Chinese open models.