Meta's Muse Spark Open lands on OpenRouter with 1M context

OpenRouter, the service that puts many models behind one API, says Meta's Muse Spark Open was available to call through it on launch day, served by four inference providers, Baseten, Fireworks, Nebius and Together (OpenRouter thread). Because the model's weights are open, outside companies like these can host it themselves rather than relying on Meta's servers; OpenRouter lists it at $1.25 per million input tokens and $4.25 per million output tokens, with a 1 million token context window and input in images, video, audio and PDFs as well as text. OpenRouter describes it as built for agentic coding and as the strongest open-weight model from a US lab. Its model page could not be read, so these specifics rest on the announcement post.

The "strongest" claim is OpenRouter's, but Artificial Analysis's independent benchmarking, published hours later and covered separately, points the same way: it describes Muse Spark Open as an 816-billion-parameter model with 42 billion active per token and scores it 39 on its Intelligence Index, the highest of any US open-weights model, ahead of NVIDIA's Nemotron 3.5 Super VL at 26 (Artificial Analysis thread). The same analysis qualifies the "built for agentic coding" framing, though: among the nine highest-scoring open-weights models, it ranks last on the index's four agentic evaluations and fourth on the other six combined.