Lev

Not yet independently verified. Interfaze’s own blog, model card and code repository are the only sources for this model; no press report or independent evaluation was found. All benchmark figures are Interfaze’s own runs. The model card says some of the training datasets carry their own terms, including non-commercial licences, which a business should review before commercial use. We will update this when it can be confirmed, and remove this note.

Lev is an open-weight decision model from Interfaze (JigsawStack, Inc.), published on 24 September 2026 with a blog post, a Hugging Face repository created the same day and the training code on GitHub a day later. It is Interfaze’s open counterpart to TypeSafe’s Jev: given a piece of context — a support ticket, an email, a JSON record — and a set of typed questions (yes or no, pick one option, or rate on a scale), it returns a probability for each option in a single forward pass and writes no text, so it cannot answer with a label outside the list it was given, though it can still choose the wrong one. It uses the same /v1/systemone request format as Jev, so code written for TypeSafe’s SDK can be pointed at a self-hosted Lev by changing the address. Technically it is a LoRA adapter of about 170 MB on Alibaba’s Qwen3.5-4B, released under Apache 2.0, and runs on one GPU or, much more slowly, on a CPU. Interfaze reports 68.9 per cent across all 13 S1Bench subsets against 76.1 per cent for Jev in its own run of both models, 0.980 on the banking77 intent set, and 414 to 654 milliseconds end to end on one H100 against about 340 for Jev’s hosted API; those are Interfaze’s own measurements and are not stored as scores. For routing, moderation, triage and intent detection at volume, the case for it is an open model a business can host itself with no per-call charge. Several unrelated projects also use the name Lev, so buyers should check they have the interfaze-ai repository.