Meta Releases Muse Spark Open with 816 Billion Parameters as Free Open-Weight Model
This article was created with the help of artificial intelligence.
Key Takeaways
- On 15 October 2026, Meta released Muse Spark Open, an AI model with 816 billion parameters specifically designed for software development, available free to developers.
- The model is based on a mixture-of-experts architecture with 816 billion parameters in total, of which 42 billion are active, and features a context window of one million tokens.
- Open-weight models such as Muse Spark Open enable developers to download model parameters and run them locally without relying on cloud services.
- Muse Spark Open is specifically optimised for running on Apple's M5 Ultra chips and requires at least one terabyte of RAM.
- According to Alexandr Wang, Meta's AI chief, the model rivals and in some cases surpasses similar products from leading Chinese developers.
Meta released Muse Spark Open on 15 October 2026, an AI model with 816 billion parameters specifically developed for software development and complex tasks. The company describes the model as the "most powerful US-based open-weight model" and makes it available to developers for free.
Open-weight architecture enables free customisation
Unlike closed models such as OpenAI's ChatGPT or Anthropic's Claude, developers working with Muse Spark Open can download the model parameters – the weights and biases – and modify them according to their own requirements. Open-weight models are AI models whose internal parameters are publicly accessible and can be operated and customised locally by users without relying on cloud services.
Meta emphasises that this approach is "cost-effective" and "offers transparency and flexibility that closed models do not enable". Despite the openness, the company states it has integrated "strong security measures" into the model.
Technical specifications and architecture
Muse Spark Open is based on a mixture-of-experts architecture with 816 billion parameters in total, of which 42 billion are active. The model is natively multimodal and features a context window of one million tokens. This technical design enables the processing of very large code repositories and complex documentation.
The model was developed by Meta Superintelligence Labs and is available through multiple API platforms, including Hugging Face. There, developers can download the model weights directly.
Focus on software development and automated tasks
According to Meta, Muse Spark Open was developed for "programming and agent-based work". The model excels at "long-term tasks, complex software development and multi-step reasoning". Meta demonstrated various use cases, including bug fixes and new features for a Mastodon client as well as video editing tasks via DaVinci Resolve, where the model reviewed and trimmed material.
Combined with Muse Code, Meta's AI-powered code terminal, the model is intended to support various automated workflows in software development.
Optimisation for Apple M5 Ultra marks unusual partnership
Muse Spark Open is specifically optimised for running on Apple's M5 Ultra chips, signalling an unusual collaboration between the two technology companies. The model can be run on Mac Studio computers with M5 Ultra chips, but requires at least one terabyte of RAM.
According to Meta, users can combine two Mac Studios with M5 Ultra chips to create a "supercomputer" on their own desk. These hardware requirements make the model interesting for professional developers and enterprises seeking to run computationally intensive AI applications locally.
Positioning against Chinese competition
Alexandr Wang, Meta's AI chief, stated the model "rivals and in some cases surpasses similar products from leading Chinese developers". So far, Chinese developers have dominated the open-weight market. Alibaba's Qwen, Moonshot AI's Kimi K3 and Z.ai's GLM models are considered "global open-weight leaders", but Meta now claims to compete at a similar level.
The release follows shortly after Meta announced a new closed model that, according to the company, "competes with the most advanced versions of ChatGPT and Claude". Meta's AI strategy has evolved from an initial focus on open models, with CEO Mark Zuckerberg recently stepping up the company's competition with OpenAI and Anthropic.
Significance for developers and enterprises
For developers in German-speaking regions, Muse Spark Open offers the opportunity to run a powerful AI model locally without relying on cloud services from US or Chinese providers. The open-weight licence enables customisations to specific requirements, such as in the areas of data protection or industry-specific compliance.
However, the high hardware requirements mean that primarily larger development teams and enterprises with appropriate infrastructure are likely to benefit from the model. Whilst smaller developers can use the model via API platforms, they lose the advantage of complete local control.