Developers post open, local alternatives to TypeSafe's Jev service
One is a standalone binary that speaks Jev's wire API and escalates to a local model, the other a 4B cross-encoder posted on Hugging Face.
Dmitri Sotnikov posted Lev on Thursday night, a decision engine that exposes the same wire API as TypeSafe's Jev service. Lev runs offline as a standalone binary, and Sotnikov's post presents it as an open alternative to the paid service.
Lev sends a question to a small encoder first, which Sotnikov's post puts at roughly 100ms on a laptop CPU, and escalates low-confidence cases to a local chat model, Qwen3.5-4B by default. The repository carries an Apache-2.0 licence, 79 commits and 10 stars.
Sotnikov's own numbers on a 144-case adversarial set put the English encoder at 61% accuracy and around 125ms, and the Qwen3.5-4B thinker at 95% and around 174ms. Nobody else has tested the figures. Lev is written in Jolt, a Clojure dialect hosted on Chez Scheme.
A second open effort is further along by one measure. AlexWortega's openjev, a cross-encoder finetuned from Qwen3.5-4B, has 556 likes on Hugging Face, and its author updated it on Thursday evening. Its card claims 0.866 accuracy on public JevBench v1.2 items for the 4B v5 checkpoint.
The openjev card is direct about what the model cannot do. It says the model has no hardening against prompt injection and reaches only 60% accuracy on shell-command safety review, and that "a guard built on this model belongs next to deterministic checks, not instead of them". The licence is MIT.