Meta on Thursday released the weights for Muse Spark Open, a natively multimodal model with 816 billion total parameters (42 billion active) and a one million-token context window.

Built for coding and agentic work, the company says it excels at long-horizon tasks, complex software engineering, and multi-step reasoning.

Meta's chief of AI Alexandr Wang called it "the most capable US open-weight model to date," and weights are downloadable from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models, with hosted inference via Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI.

The twist builders are talking about is where it runs. Muse Spark Open is the first model in its class to run locally on a desktop using two Apple M5 Ultra Mac Studios with the latest MLX framework.

Apple's unified memory lets each Mac Studio's integrated GPU use up to 512 GB, so two machines hold the entire model and, per Meta, "behave like one supercomputer with a terabyte of unified memory."

Zuckerberg credited work with Apple's John Ternus, and The Wall Street Journal put the hardware cost at about $40,000 — computing power that could cost thousands of dollars monthly to rent.

Meta says it tested agentic workflows running entirely on-device, using Muse Code powered by Muse Spark Open to ship real iOS changes to Ice Cubes, an open-source Mastodon client — bug fixes, new features, and a full build — and to drive DaVinci Resolve end to end, with the model reviewing footage frame by frame to find the strongest moments and crop them vertically.

TechCrunch notes the local requirement is steep, at least 1 terabyte of RAM, and reports the model was also paired with Muse Code in agentic coding workflows. Meta claims in evaluations that the model rivals, and in some cases outperforms, Chinese open-weight equivalents.

The release lands in a market where Chinese models dominate.

Rest of World reports that four Chinese models — DeepSeek, z.ai, Xiaomi, and Tencent — held four of the top five spots on the OpenRouter leaderboard last week, and that LiveBench recently showed top Chinese models trailing US rivals by just 3%.

Stanford's Graham Webster said at the Rest of World event that if you need relatively frontier capability, "right now that's Chinese open models only," while Trillium Labs' Nathan Lambert wrote that Reflection, like Nvidia and Thinking Machines, released its strongest models and "come up behind Chinese counterparts."

Meta isn't alone this week.

Reflection released the weights for Beam, its first open-weight model, under an Apache 2.0 license — a 501B-parameter mixture-of-experts model (23B active) with a one million-token context window that unite.ai reports scored 90.5% on GPQA Diamond and 80.9% on SWE-bench Verified.

StepFun also open-sourced Step 5 Preview, a 600B-parameter sparse MoE model that 36Kr reports ranked second globally on OpenRouter's weekly token usage list and costs about one-eighth as much per task as Claude Opus 5.

Rest of World notes Chinese open models can run tasks 50% to 90% cheaper than leading American frontier models.

Zuckerberg told investors on the Q2 call that Meta would likely return to open models and mix open and closed releases, and TIKR notes the release widens the models' reach without generating revenue on its own — Meta is also exploring an API and business agents.

Meta says Muse Spark Open carries additional safety post-training and met the thresholds of its Superintelligence Scaling Framework, with limited commercial terms under the Muse Model License.

The open question, per Rest of World, is monetization and whether openness can win markets: whether builders pick a US-made open model at parity, or stick with the cheaper, entrenched Chinese alternatives.