Former OpenAI researcher Diogo Almeida has unveiled the first public AI model from TypeSafe AI. The project is named Jev and was developed with a different operational concept compared to traditional chat models. Developers believe that language systems that sequentially generate text token by token are not always suitable for software tasks. Instead of detailed responses, Jev returns structured results, probabilities, and confidence indicators. Several solutions can be processed in parallel.
TypeSafe AI classifies Jev as a model belonging to the 'System One' category. This approach is linked to the division of decision-making processes into quick and more prolonged ones. Developers propose using the model directly within programs: the code handles operations, while Jev is integrated where a choice between options or an assessment of an intermediate result is required. The company claims that the model can operate 20–200 times faster than traditional chat models, and its use is 40–400 times cheaper. The cost of input tokens is stated at $0.042 per million, and output tokens, according to the presented terms, are not charged.
For training Jev, TypeSafe AI used its proprietary method called RLCD. A detailed description of the technology is not yet available in the public announcement. The company was founded in 2024 and focuses on developing AI systems for integration into software. According to Almeida, by March 2026, the startup had attracted over $35 million. Access to Jev is being provided gradually: the first invitations to engineering teams began in January through a waiting list, and now some access is distributed via the Discord community. However, independent performance tests and comparisons with other AI models have not yet been published, so the claimed characteristics remain the company's own assessments.
Related articles
