Von: Open-source 395M "System One" model replacing TypeSafe's JEV, runs on CPU

A Reddit user independently benchmarked three new "typed decision model" projects on their own private repo's commit history instead of trusting readme scores. Hosted proprietary Jev won every task, stock Qwen-based Semif placed second on CPU, and purpose-trained Von 1.0 performed at or below trivial baselines. The poster emphasizes the method (labels from conventional-commit prefixes, free baselines scored alongside) transfers in an afternoon.

Social·Community

JEV explained: TypeSafe's decision model reads once and scores all allowed answers — no hidden 'none of these' option

A Diamant AI newsletter post explains the internals of JEV, a decision model from TypeSafe. Unlike an LLM loop, JEV reads once and scores every allowed answer in parallel. Its output probabilities must land on one of the supplied options, so practitioners should always include an escape answer.

A Diamant AI newsletter post explains the internals of JEV, a decision model from TypeSafe. Unlike an LLM loop, JEV reads once and scores every allowed answer in parallel. Its output probabilities must land on one of the supplied options, so practitioners should always include an escape answer.

A paper benchmarks JEV, a commercial 'System One' model from TypeSafe, against GPT-6 Luna and Qwen3.8-27B across seven political science replication tasks. JEV matches or comes close to both LLMs in accuracy and shows better calibration than GPT-6 Luna's token probabilities, but shows no cost advantage over GPT-6 Luna at OpenAI's batch prices.

Releases·Model Release

LibertAI releases Deem, an open-weight Jev alternative built on Qwen3.5-9B

LibertAI released Deem, an open-weight model positioned as an alternative to Jev, built on Qwen3.5-9B with a smaller 0.8B variant. It trails Jev on hard benchmarks but arrives early in a new model category, and the small version runs on CPU for local tinkering.

LibertAI released Deem, an open-weight model positioned as an alternative to Jev, built on Qwen3.5-9B with a smaller 0.8B variant. It trails Jev on hard benchmarks but arrives early in a new model category, and the small version runs on CPU for local tinkering.

Products·launched

Idea scorer that tests TypeSafe's Jev model, a non-generative classifier, via ~10 parallel questions

A side project where users enter a project idea and receive a score after answering around 10 classification questions, all run in parallel. It tests TypeSafe's Jev model, which outputs scores instead of generating text. It aims to help people quickly validate ideas faster than a sequential LLM evaluation would.

A side project where users enter a project idea and receive a score after answering around 10 classification questions, all run in parallel. It tests TypeSafe's Jev model, which outputs scores instead of generating text. It aims to help people quickly validate ideas faster than a sequential LLM evaluation would.