Zuck’s Muse to Spark joy with open weights release ‘soon’
Meta CEO Mark Zuckerberg has said the company will "soon" release an open-weights version of its flagship AI model, Muse Spark, though no firm date was given. In the meantime, an updated version, Muse Spark 1.3, has gone live on Meta's API service and its Muse Code CLI, offering efficiency and reliability improvements aimed particularly at users running the model with AI agents and code assistants. The release keeps up the pace of refinements Meta has made since launching Muse Spark in April, as it seeks to stay competitive with rivals and reassure investors about its heavy AI spending.
The new version is designed to perform better on long, multi-step tasks and to be more realistic about its own limitations, asking clarifying questions, seeking human help when stuck, and confirming before taking significant actions, rather than wasting tokens on dead ends. Meta says this also reduces the number of turns and tokens needed to complete tasks, cutting costs further. Independent benchmarking by Artificial Analysis reportedly shows a four-point jump in overall intelligence scores, putting Muse Spark 1.3 roughly level with GPT 5.6 Sol, Claude Opus 5 and Grok 4.6 High, while a discounted "contributor tier" offers pricing as low as $0.002 per million cached input tokens for users willing to accept lower rate limits and data-sharing terms.
In short
- Meta will soon open-source Muse Spark's model weights, Zuckerberg says
- Muse Spark 1.3 launched now, more efficient and less error-prone
- Benchmarks put it level with GPT 5.6 Sol, Claude Opus 5, Grok 4.6
Steelman: both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
An open-weights release could give developers and researchers more ability to inspect, adapt and run Muse Spark for their own needs, while Meta’s efficiency improvements and low-cost contributor tier may make capable AI tools more accessible. Releasing weights could also encourage experimentation and competition beyond Meta’s own services.
The case against
Making a powerful model’s weights available can make it harder to control how the system is modified or used, and could increase risks if safeguards are removed or adapted for harmful purposes. The contributor tier’s data-sharing terms and lower rate limits may also leave users weighing affordability against privacy and reliability, while benchmark gains alone do not establish how well the model performs in practice.
Explained simply
Meta is Mark Zuckerberg’s technology company, best known for Facebook, Instagram and WhatsApp. It is also investing heavily in artificial intelligence, including Muse Spark, its flagship AI model. AI models are software trained to respond to questions and carry out tasks such as writing or coding.
An open-weights release would let other people download the model’s learned parameters and run or adapt it themselves, subject to the terms Meta sets. This differs from using a model only through a company’s online service, though “open weights” does not necessarily mean all the training data or development methods are shared.
AI companies compete on how capable, dependable and affordable their models are, especially for coding and systems that can take several steps on a user’s behalf. Meta’s releases matter to developers choosing tools and to investors watching whether its costly AI investments can keep pace with rivals.