Meta releases Muse Spark Open, an 816B-parameter open-weight model with 1M-token context

Meta released Muse Spark Open, an open-weight model with 816B total parameters and 42B active parameters, featuring a 1M-token context window. The model is designed for coding and agentic work and is natively multimodal.

According to the model card on Hugging Face, Muse Spark Open underwent additional safety post-training to enhance robustness against misuse when deployed as a standalone model without system guardrails. Meta recommends deploying the model as part of a broader AI system with use-case-specific guardrails.

Performance-wise, Muse Spark Open is described as falling above GLM 5.2, near DeepSeek V4.1 Flash, and behind GLM 5.3 or Kimi K3. The model’s capabilities align with the increasing demand for open post-training economies and enterprise adoption of AI systems.

Meta also released closed frontier models Muse Wonder and Muse Wonder Full, positioning the company as the only American lab with both open and closed frontier models. On traditional benchmarks, these closed models are reported as very strong, only behind Anthropic’s Opus 5.5 and similar to OpenAI’s GPT-6-Astra or Google’s Gemini-4-Argon.

The release of Muse Spark Open comes alongside Reflection’s Beam model and Mistral’s Mistral Large 4, which collectively reduce the performance gap between American and Chinese open-weight models by roughly half. Without these releases, Chinese companies would have held a commanding lead with over 20 open-weight models outperforming the best Western models.

Muse Spark Open is distributed under a custom license that includes restrictive terms such as a requirement to display 'Built with Muse' branding on products using the model. The license also includes a revenue-gating clause: companies generating more than $20M per month across their entire business deploying the model require a separate license from Meta. Additionally, the license mandates compliance with an acceptable use policy, which the author argues makes the model less open than traditional open-source software.

The author notes that while the license’s legal ambiguity may deter some users, many organizations are likely to ignore it or proceed despite the restrictions. The license’s complexity and untested legal status introduce strategic risks for labs aiming to compete and regain adoption in the open-model ecosystem.

Meta’s recent model releases follow a strategic redirection under CEO Mark Zuckerberg’s vision for a high-ownership, high-compute-per-researcher lab, which the author suggests may pay off despite past criticism. The company’s dual approach—releasing both open and closed frontier models—reflects a unique strategy in the current competitive landscape.