TypeSafe AI @typesafeai's Jev is now fully open, and the Cognition team has responded with an open-source Kev. The signal from the AI world over the past two days is clear: we no longer need to burn expensive compute on a “yes or no” judgment.

I have an automated workflow for processing massive volumes of leads. I used to rely on GPT-4o to screen 700 potential customers—not only was it costly, I also had to worry about it occasionally hallucinating a justification.
Yesterday, I tried modifying the workflow: I had Jev score the matches, assigning each lead a confidence score and immediately filtering out those below the threshold. The entire batch finished in 40 seconds and cost less than ten cents.
The core idea behind this kind of “decision model” is that it does only one thing: select the optimal solution from the options you provide, or score your statements. It does not write poetry or chat, so it can deliver 20x or even 200x speed improvements.
And the barrier to entry has now truly been lowered to the floor. Kev makes local deployment readily accessible: training a 4B model on a single H100 takes just 40 minutes. This means you can fine-tune a dedicated decision-making model for your specific business data in minutes. Kev-8B’s performance on out-of-domain data is already approaching Jev’s, showing that this path is entirely viable.
Try Jev or Kev and replace your AI routing, intent recognition, and security gateways with this kind of specialized decision engine. When you see your agent’s response time drop from seconds to milliseconds, you’ll realize that this is the real efficiency dividend.