TypeSafe AI's Jev model passes one trillion tokens in three days on Modal
TypeSafe AI's first model, Jev, scaled to more than one trillion tokens in three days after its public launch, running on Modal's serverless infrastructure, according to a Modal case study. Jev is not a large language model but what the company calls a System One Model, built for fast, structured decisions.
- TypeSafe AI's Jev served over one trillion tokens within three days of launch on Modal, according to Modal.
- Jev is a non-autoregressive System One Model that returns typed, structured values with probabilities instead of text.
- TypeSafe trained it with a method called Reinforcement Learning for Calibrated Decisions (RLCD).
- One week after launch, Jev topped OpenRouter's leaderboard for 1K-10K token contexts with 15.1% of requests.
A decision model, not a text generator
TypeSafe AI describes its approach with the manifesto "Build Prod, Not God." Its first model, Jev, is designed to deliver fast, structured decisions with frontier intelligence. According to the case study, Jev is a System One Model, a reference to fast, intuitive thinking. It reaches similar levels of intelligence on System One tasks when compared to frontier LLMs, but it focuses on decisions rather than generalized intelligence.
To build it, TypeSafe created a custom stack focused on automation. The stack includes a new model architecture, a parallel sampler for maximum efficiency, and a training method the team calls Reinforcement Learning for Calibrated Decisions (RLCD).
Unlike LLMs, Jev is not autoregressive and does not generate text. It takes a state and a set of questions as input and returns typed, structured values with probabilities. Because it produces outputs in parallel rather than sequentially, Modal says it gives low-latency responses suited to large-scale automation, fast browser use and generative user interfaces.
How a request looks
The case study shows a customer support ticket as an example. The input is a message from a user who has spent three days trying to connect a Stripe account. The request asks which team should handle the ticket and defines a choice primitive with three options: billing, technical and sales, each with a short description.
Jev responds with a structured answer. In the example, it picks technical with a confidence of 0.78, and lists probabilities of 0.85 for technical, 0.0 for sales and 0.15 for billing. The user defines the options, and the model selects from them.
Demand after launch
Jev launched publicly on a Thursday. One week later, according to Modal, it topped the OpenRouter leaderboard for contexts between 1K and 10K tokens, handling 15.1% of requests. The case study says this is only a slice of the demand TypeSafe handles, which the team now measures "in the trillions" of tokens.
To serve Jev at the price-performance frontier, TypeSafe needed full control of its stack, from networking and routing to autoscaling for unpredictable load. It wanted that control without managing its own infrastructure.
What Modal provides
Modal presents its platform as suited to bursty, high-concurrency workloads where every millisecond counts. Instead of guessing capacity needs or overpaying for a compute reservation, AI labs can use its serverless platform to train, serve and scale types of models including LLMs, System One Models and world models.
TypeSafe CTO Erik Gafni is quoted in the case study: "Modal served 1 trillion tokens for Jev within three days of launch. Their team was proactive and responsive as we scaled to meet unprecedented demand, and we could stay focused on shipping instead of managing infrastructure."
The case study is published by Modal, so the performance and demand figures come from the vendor and its customer rather than an independent evaluation. It positions Jev's launch as an example of what the platform was built for: bringing any model and any serving code, and scaling to trillions of tokens per day.
