kettei
Sqwish Labs' decision-model family returning answer probabilities in one pass, with open 3.7B and 900M models and a fine-tuning API (OneHot).
Overview
Sqwish Labs announced the kettei model family on October 14, 2026. Each model takes a context, questions and allowed answers and returns a probability per answer without generating text. kettei core (3.7B) and dot (900M) are text-only open-source models built on the Institute of Foundation Models' K2-Horizon base; fat and chub also accept images and are served through OneHot, Sqwish's inference and fine-tuning platform. Sqwish quotes about 8 ms GPU time for core and 26 ms for fat per short request, with training across 78 languages.
This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.
Key features
- Four models: fat and chub (text+image), core (3.7B) and dot (900M) (text-only)
- core and dot are open source; fat and chub are OneHot-only
- Inference billed on input tokens only; no charge for generated output
- Fine-tuning live for chub and core (about 15 minutes); fat fine-tuning in early preview
- 78 languages and 24 writing systems; US and EU regional deployments
Use cases
- Classification, routing and moderation decisions in agent pipelines
- Low-latency scoring where text generation is unnecessary
- Fine-tuning small decision models on proprietary labels
Pricing
Input tokens only; fine-tuned inference costs 2x the base rate; training billed per million tokens with a $2.50 minimum per job. Base per-token rate not published.
Pricing reflects the source record and may change. Confirm current terms with the provider.
Dates and sources
- Released
- 14 Oct 2026
- Record updated
- 15 Oct 2026
- Record last checked
- 15 Oct 2026