Kev adds 0.6B, 4B and 8B open-source decision models based on Qwen3

@jaredpalmerUPDATE: Kev-0.6B, 4B, and 8B are now available. Kev is a family of small open source Jev-like decision models you can train and run yourself. This new family is based on Qwen3 using the same LoRA + small pointer head technique as before, but scaled up. Out of domain, on data Kev never trained on: Kev-8B 79.6%, Jev 85.7%. • Drop-in TypeSafe System One API; their SDK works with one `base_url` change • Kev-4B serves on a 32 GB Mac in bf16: ~300 ms for five questions, ~40 ms on an H100 • Repeated documents hit a KV cache: 2-2.5x faster • Apache 2.0 License. Kev-4B trains in 40 minutes on one H100. Kev-8B in 83 minutes. Code, weights, evals: https://github.com/jaredpalmer/kev

@jaredpalmerWhile Jev still lead most benchmarks, Kev was able to outperform Jev on some logic recognition tasks. Out of the family, the 4B model is the most interesting for sure.

@jaredpalmerRelative to my initial 0.5B Qwen2.5 prototype, what moved the needle? • Backbone 0.6B to 4B: +14 to +19 pp • Learning rate 2e-4 to 5e-5: +4.7 pp • Random rule structures in the synthetic data: +1.5 to +3 pp • 4B to 8B: +0.5 to +2 pp The learning-rate result was interesting. The base model already knows things (MMLU, paraphrase). The default fine-tuning recipe was erasing them. Cutting the learning rate 4x recovered most of that. So it turned out that more public training data raised in-distribution accuracy and lowered out-of-domain accuracy 🙃

@jaredpalmerWhere Kev still loses to Jev: knowledge (MMLU 70 vs 90) and day-precision date arithmetic (60 vs 93). In practice, I expect Kev to be useful for teams to finetune their own Jev-like models on their own data to bring their inference cost even closer to zero. After hill climbing all weekend, my entire bill on @modal was just $228. Most training runs cost between $10 and $35.

@jaredpalmerThe entire Kev family is 100% open source. Code, evals, etc on GitHub: https://github.com/jaredpalmer/kev/tree/main Weights, model cards, datasets on HF: https://huggingface.co/collections/jaredpalmer/kev

@ekzhang1@jaredpalmer Did you try evaling it on MMLU-Pro (not MMLU, saturated) and also update backbone to Qwen3.5 or higher? The qwen3 models are pretty out of date

@jaredpalmerI can't run on it as-is btw. 3 of every 4 layers are Gated DeltaNet (linear attention, recurrent). Kev's core trick (packing the document and all questions into one sequence with a block-causal mask so questions can't see each other) only controls the attention layers. Through the recurrent layers, question 2 would see question 1. This breaks the current isolation guarantee in Kev

@ChengleiSiRT @jaredpalmer: Kev-0.5B: A tiny open source Jev-like decision model with a TypeSafe-compatible API based on Qwen2.5-0.5B that you can tra…

Kev adds 0.6B, 4B and 8B open-source decision models based on Qwen3

TLDR

Combined views

Kev adds 0.6B, 4B and 8B open-source decision models based on Qwen3

8 Sources

Combined views

Sentiment

Summary

TLDR

Sentiment

Sentiment

Summary

Sentiment

8 Sources