It seems that Jev is also learning using synthetic data and labels from a large-scale model
It seems that Jev is also learning using synthetic data and labels from a large-scale model, so I distilled and learned the judgments of DeepSeek V4 Flash with a weight of 157GB over 26 hours using DGX Spark, and transferred it to the 4B model. It is 1/20th the size and exceeds the teacher's immediate response mode, approximately 22ms per judgment. I never thought I could do this much with Local LLM alone.
They gave Astra, DeepSeek 4.1, Fable 5, and Qwen 3.8 the same prompt for the boat game.
The comparisons are really terrible 🤯
Since I've got access to Jev from typesafeai , I've added to one of my CRE applications
You ask something → DeepSeek understands it and fills in the inputs → Jev confirms which specialist this belongs to → the local engine runs → DeepSeek gets a small facts packet of
Who still doubts that JEV is useless? Actual testing of JEV’s real application scenarios. I continued to use JEV and DeepSeek to run the same task.
Simulate the company's real massive inbox, divert 300 messy letters (business procurement, online failures, compliance and legal affairs, promotional waste) to 15 business departments in real time, and label them with intent Left: TypeSafe Jev (jev-latest)
You have reached the end of the archive
All of deepseek