Meta's 816B open-weight launch: day one, everywhere

Meta released a new open-weight model on October 15, and the interesting part is not the parameter count. Muse Spark Open, from Meta Superintelligence Labs, is a mixture-of-experts system with 816 billion total parameters and 42 billion active per token, a 1-million-token context window, and native multimodal input (text, images, video, audio) with text output. It is aimed squarely at coding and long-horizon agentic tasks. The model card shows 256 experts per layer with 8 selected per token, the same routing philosophy that has powered the largest open releases out of China over the past 18 months.

The specs are Meta's own. The launch footprint is verifiable and much more unusual. Muse Spark Open went live simultaneously on seven hosted-inference providers (Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, Together AI) and four weight-distribution channels (Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, Microsoft Foundry Models). Microsoft confirmed its side directly through the Azure AI Foundry blog, and Intel announced day-0 support through its own newsroom. Most open-weight releases drop weights on Hugging Face and wait weeks for infrastructure providers to catch up. Meta lined up eleven distribution partners before the announcement went out. That is a product launch with an enterprise sales motion behind it, not a research drop.

Meta's marketing line, that this is "the most capable US open-weight model to date," is a direct jab at the Chinese labs that have dominated open-weight leaderboards for the past year. Treat it as Meta's claim. No independently reproduced benchmarks against DeepSeek, Qwen, or GLM were published alongside the release. The capability ranking is something to watch, not something confirmed. Correction (2026-10-16): the first independent numbers have since landed; see the correction box below. They give Meta's US ranking independent backing but keep the global caution intact.

Two claims deserve separate handling. Meta demonstrated the model running on two Apple M5 Ultra Mac Studios through Apple's MLX framework, which, if it holds up, puts an 816B-parameter model under a desk instead of in a data-center GPU cluster. That is a Meta demonstration claim, not an independent test. And the "two-tier" framing: Meta's Chief AI Officer has said the company will generally not open-source its most powerful models. Muse Spark Open is reportedly not Meta's top model; the closed Muse Wonder, which we covered October 13, keeps that distinction. The open release buys developer goodwill and fine-tuning ecosystems; the closed flagship keeps the best capabilities monetised.

The operator decision: if you are evaluating open-weight models for agentic coding work, Muse Spark Open enters the shortlist on distribution breadth and context size alone. Day-one availability across this many providers means pricing competition starts immediately, and the 1M-token window means whole repositories in one pass. But run your own evals before trusting the "most capable" framing, and watch for independent benchmark numbers before treating it as a DeepSeek or GLM replacement. No hands-on testing was performed for this briefing.

Related: our Muse Wonder briefing covers Meta's closed top tier, and Claude Haiku 5.5 subagent economics is the cost-side lens for agentic deployments.