Reverse-engineering Jev: likely a Qwen 3.5-based 35B-A3B decision model

Taenin · reddit · 2026-10-27

Oumi researchers reverse-engineered TypeSafe's decision model Jev. Key findings:

Base: likely Qwen 3.5/3.6, based on obscure-vs-famous knowledge comparisons, name probing, and sensitive-question response patterns
Scale: likely 35B-A3B (3B active), inferred from benchmark performance and serving profiles
Tokenizer: probably a merge of Qwen and GPT-OSS tokenizers
Training: the claimed RLCD is likely standard soft-label cross-entropy on synthetic data (distillation-like), not PPO/GRPO
Synthetic data: a mix of reasoning-model teacher outputs and programmatic answers to code-generated questions
Serving: likely tree batching (all questions packed into one request with per-question attention masks) instead of the standard vLLM path; long-document performance suggests sparse attention or two-GPU split prefill, probably on B200s or above

More from Models