Sqwish launches kettei decision models on OneHot
On 14 Oct 2026 Sqwish Labs launched OneHot, an inference and fine-tuning platform, with a family of decision models called kettei: fat and chub (text and images) and core (text only), plus a 900M dot that its launch post names. Per the launch post, the models take a context and a list of allowed answers and return one probability per answer in a single pass. The API has a native /v1/decide route and a /v1/systemone route that Jev clients can point at. Per the docs, input is billed at $0.06/M for fat, $0.04/M for chub and $0.02/M for core, with output free; new accounts may start on a waitlist.
Independent results. On Benchmark Heaven’s JevBench API board (v1.7.46, 17 Oct) kettei-fat is #1 with a composite of 77.8, ahead of d1-ultra (76.1) and Sage 1.3.0 (74.0); kettei-chub is #4 at 73.5; Jev 1.13.0 is 71.5. On ImageJevBench v0.3.0 kettei-fat is also #1. Those boards are separate and are not mixed with others.
What is unverified. Sqwish’s own figures (93.1% on the 231 public JevBench items, 97.5% on its own held-out image set, a Decision Index run on a 12,000-request sample) are vendor claims. The post calls core (3.7B) and dot (900M) open source, but on 17 Oct the Sqwish Hugging Face organisation lists no models, we found no sqwish GitHub organisation, the docs treat core as an API model and dot is not in the docs or price list. Until weights and a licence appear we treat the family as a closed API in beta. The page scores the fat size. See the kettei page.