THE ESSENTIALS IN 30 SECONDS

  • Muse Spark Open has 816B total parameters, 42B active per token, and a 1M-token context window.
  • Meta's own figures put it ahead on tool-use benchmarks such as MCP-Atlas but behind Kimi K3 and GLM 5.3 on Terminal-Bench 2.1.
  • The Muse Model License requires a separate agreement for services run by companies above $20 million in annual revenue.

Meta Superintelligence Labs released Muse Spark Open on October 15, an open-weight version of its Muse Spark model that the company describes as the most capable US open-weight model to date. The release is pitched at coding and agentic work, including long-horizon software tasks and multi-step reasoning.

What was released

According to the model card, Muse Spark Open is a mixture-of-experts transformer with 816B total parameters and 42B active per token (62 layers, 256 experts, 8 used per token). It accepts text, image, video and audio input, produces text, and supports a 1M-token context at inference. Weights come in BF16 and NVFP4; Meta estimates self-hosting needs roughly 700GB of GPU memory in NVFP4 and about 2TB in BF16.

Beyond Hugging Face, the weights are listed on Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry, and hosted APIs are offered by providers including Fireworks AI, Groq, Together AI and OpenRouter. Meta also says the model can run across two linked Mac Studios with M5 Ultra chips using Apple's MLX framework.

How it compares

Meta's published numbers are mixed. The model scores 89.2 on MCP-Atlas, a tool-use benchmark, ahead of Qwen3.8 at 84.5. On Terminal-Bench 2.1, however, it reaches 76.4 against 85.0 for Kimi K3, and its Artificial Analysis Intelligence Index score of 39 trails GLM 5.3 and Kimi K3. The New Stack notes that Meta's coding gap with leading Chinese open models, acknowledged when the closed Spark launched in April, has not closed much. All figures are vendor-reported.

Licence and safety caveats

The weights are gated and released under the Muse Model License rather than an open-source licence. Products must display "Built with Muse Spark Open", and companies whose services exceed $20 million in gross revenue over 12 months must notify Meta and sign a separate agreement before offering the model to third parties. Internal use and research are excluded from that requirement.

Meta says it applied extra safety post-training because the weights are public, and reports high refusal rates on its internal cyber and biology tests. It also warns those results apply only to the released checkpoint: fine-tuned derivatives may behave differently. A full evaluation report has not yet been published.