Meta releases Muse Spark Open weights: 816B-A42B multimodal model with 1M context, built for coding and agents

Why it matters

【Key points】

On October 15 Meta released the weights of Muse Spark Open, which it calls the most capable US open-weight model to date. It is natively multimodal (text, image, video, audio) with 816B total and 42B active parameters and a 1M-token context window, built for coding and agentic work: long-horizon tasks, complex software engineering and multi-step reasoning.

Weights are on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry, with an NVFP4 variant; Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter and Together serve it via API from day one. With Apple's MLX the full model runs on two M5 Ultra Mac Studios (up to 512GB unified memory each), and SGLang, vLLM, TensorRT-LLM, Transformers, llama.cpp and Unsloth support it. The license is the Muse Model License with limited commercial terms that Meta says let many developers use it for free. Meta says the model met the release thresholds of its Superintelligence Scaling Framework; a full evaluation report is coming.

【Verifiable】

Meta research blog: https://research.meta.ai/blog/muse-spark-open

Hugging Face collection: https://huggingface.co/collections/meta-models/muse-spark-open

@AIatMeta launch post: https://x.com/AIatMeta/status/2110779754274292212

【Our read】

1. A near-trillion-parameter open-weight model from a US lab puts direct pressure on DeepSeek, Kimi, Qwen and StepFun; open weights are no longer a mostly Chinese race.

2. "Most capable" is Meta's own claim, shown only in blog charts; wait for third-party evals and the full safety report.

3. Open weights is not open source: read the Muse Model License's commercial limits before production use.

4. Running on two Mac Studios matters more as a signal than for adoption; the hardware bar is still high, but frontier-class agent models on a desktop are now real.