The new Jev framework from TypeSafe AI returns probabilities and scores instead of generating text to optimize artificial intelligence for decision-making and composable intelligence. This "decision model" operates up to 200x faster and 400x cheaper than traditional large language models. It is currently available for developers through a dedicated Jev Playground and the Venice API.
This transition toward "System One" models allows for the rapid processing of large datasets, such as the re-scoring of 4,005 FOMC public speeches in the Fedlock corpus. While proponents describe the approach as intelligence "too cheap to meter," some critics argue the launch was misleading and that the tool acts as a fast classifier rather than a new frontier in AI intelligence.
Key sources
- SOURCE@mtslive“Diogo Almeida claims Jev is up to 200x faster and 400x cheaper than LLMs”x.com
- SUPPORT@antonosika“you get a jev API free via lovable when you build an agent/web app with it this week”x.com
- SUPPORT@thestalwart“re-score the entire Fedlock corpus. 4005 FOMC public speeches and statements since the mid 90s”x.com
- SUPPORT@simonw“new category of system one aka decision models”x.com
- SOURCE@hwchase17“hosting SemIf and offering it for free through LangSmith Gateway for the next week”x.com
- SOURCE@hwchase17“Score every production trace instead of a sample”x.com
- SUPPORT@hwchase17“semif is pretty close to jev (75.4 to 74.7) on JevBench”x.com
- SOURCE@hacubu“Decision models are becoming a first-class part of the LangSmith Gateway!”x.com