Open source Kev model released by Cognition team
For developers, Kev offers a low-cost, open-source alternative to proprietary decision models like Jev, with a small footprint and fast inference times.
The Cognition team has released an open-source Jev-like decision model called Kev.
It's built on the Qwen3 base model with LoRA adapters and a pointer head, available in 0.6B, 4B, and 8B parameter sizes.
The Kev-8B model achieved an accuracy of 79.6% on out-of-domain data, compared to 85.
7% for the proprietary Jev model.It can be trained and deployed locally, with the 4B version running on a 32GB Mac in about 300ms for five questions.
Cognition 团队 @jaredpalmer 开源发布了 Jev-like 小型决策模型 Kev,Qwen3 底座 + LoRA 适配器 + 指针头,可自行训练、本地部署,有 0.6B / 4B / 8B 三种参数量 随着各家 Jev-like 模型的发布,Qwen 小模型的含金量还在上升! Kev 模型: github.com/jaredpalmer/kev 能力基准:在从未训练过的域外数据上,Kev-8B 准确率 79.6%,托管专有模型 Jev 为 85.7%;开源自托管方案与商用服务还差约 6 个百分点,但已进入可用区间。 四个关键卖点: · 即插即用:完全兼容 TypeSafe System One API,官方 Python SDK 只改一个 base_url 就能指向本地服务器 · 消费级硬件可跑:Kev-4B 在 32GB Mac 上以 bf16 服务,五个问题约 300ms;H100 上约 40ms · KV cache 加速:重复文档的前缀缓存命中后快 2–2.5 倍 · 训练成本极低:单张 H100,Kev-4B 训练 40 分钟,Kev-8B 仅 83 分钟 Jared Palmer @jaredpalmer UPDATE: 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: github.com/jaredpalmer/kev 🔗 View Quoted Tweet📊 3 ⚡
This report is categorized under AITOP's "Models" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include Kev, Qwen3, decision models, open-source models.
Source attribution: curated from shao__meng (@shao__meng). Cross-source corroboration: 2 independent sources have covered this development (including arXiv: OpenAI), confirming industry-wide signal strength.
- arXiv: OpenAI09-21 17:57Source