Jev introduces a new shape of LLM
5 hours ago
- TypeSafe AI unveiled Jev, a new 'System One' decision model that outputs floating point numbers for categories, yes/no answers, and ratings with confidence scores instead of generating text.
- Jev is extremely fast and cheap: input costs $0.042 per million tokens with free output, making it cheaper than GPT-5 Nano.
- It answers three question types: Yes/No (Noul), Choice (select one from options with probability distribution), and Score (rate on a numeric scale with descriptions).
- Applications include classification tasks like spam detection, labeling, prioritization, and search reranking (e.g., scoring 100 BM25 results for relevance).
- Concerns include its black-box nature, lack of explainability, and potential bias; rigorous evals are critical before deployment.
- The community has built unconventional tools using Jev, such as a terrible chat model (jevchat), a left-pad implementation (jev-leftpad), and a 2048 game player (jev-2048).
- Open-weight recreations like Kev (based on Qwen 3.5) and a JevBench benchmark have emerged within a week of release.