← sources

How to Build a Decision Model Like Jev

source: reddit (reddit:LocalLLaMA)published 2026-10-27 15:00ingested 2026-10-28 02:38

retained text

Full disclosure: I'm one of the authors of this blog, and work at Oumi.

I thought folks here would find our investigation interesting, especially anyone interested in building local decision models.

We ran several experiments to investigate which model Jev may be based on and how it might be trained and served. The TypeSafe team has publicly mentioned that they are not interested in pretraining a model from scratch, so Jev is almost certainly built from a recently-released publicly available model -- we were curious if we could identify exactly which one! We don't have any insider knowledge, so these are our best guesses based on the results:

\- Jev is likely built on Qwen 3.5 or 3.6, based on obscure / famous knowledge comparisons, name probing, and sensitive question responses  
\- Jev is likely a 35B-A3B model, based on the benchmark performance and serving profiles we ran (see in particular the Rare Facts analysis and curves)  
\- Its tokenizer is likely a merge of Qwen and GPT-OSS  
\- Its training methodology (RLCD) is likely standard soft-label cross entropy from synthetic data. It's likely being framed as RL because of a parallel with RLHF (for which the human feedback part was the unique part), but it likely isn't using PPO or GRPO or anything like that.  
\- This synthetic data is a mix of teacher-taught from reasoning models and programmatic answers to code-generated questions.  
\- Serving is likely done with tree batching (pack all questions + 1 state into single request with per-question attention mask) rather than the standard serving path on vLLM  
\- Based on long document performance, they're likely using a form of sparse attention or serving on two GPUs and splitting the prefill between them. Most likely on B200 GPUs or above.

Our nitty gritty analysis is in the article for those curious.

Hope you folks enjoy!

document fetched 2026-10-28 00:42 · canonical https://reddit.com/r/LocalLLaMA/comments/1xguv3i/how_to_build_a_decision_model_like_jev