kettei
Quadrant scores
See the full quadrantScored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.
- Maturity5.4/10
Paid OneHot API with docs, Python SDK, agent plugin, fine-tuning and a /v1/systemone Jev-compatible route. New accounts may start on a waitlist, no SLA, no GA claim. The 'open source' core and dot weights could not be found, so they do not count.
- Capability8.1/10
Binary / single / ordinal, text and images (fat, chub). #1 on the JevBench API board (fat 77.8, calibration axis 93, board p50 0.20 s) and #1 on ImageJevBench (75.9). No ECE or method published. Scores are for kettei-fat; not on DI 0.3 or AA.
- Adoption10/100
Launch blog on 14 Oct, small X pickup (a JevBench-update post with 23 likes, a handful of others), JevBench and ImageJevBench listings. Three days old, no gateway listings, integrations or named users.
Vendor claims
- kettei fat leads every listed decision model on the public Decision Index (Sqwish's run with the maintainer's kit on the stratified 12,000-request sample) and on the 231 public JevBench items: 93.1% at a median 160 ms, run from the UK against each provider's public API[Vendor claim — not independently verified]
- 97.5% (fat) and 96.9% (chub) on Sqwish's own held-out set of 5,000 image questions, against 73.5% for Cloudflare Clef and 79–86% for frontier LLMs[Vendor claim — not independently verified]
- kettei core (3.7B, built on IFM K2-Horizon-3.7B) and kettei dot (900M, K2-Horizon-0.9B) are open source and lead every model of their size[Vendor claim — not independently verified]
- With kettei fat in front of Gemini 3.8 Flash, the pair matches Gemini's accuracy at 44% of its cost, and four in five requests never reach Gemini[Vendor claim — not independently verified]
News covering this model
publicado (OK do Fabricio 17/10 BRT).
Read this first.
- One family, four names. The blog lists kettei fat, chub, core and dot. The API serves fat, chub and core. dot (900M) is not in the docs, the price list or
/v1/modelsat all.- “Open source” is not verified. The blog says core (3.7B) and dot (900M) are open-sourced. As of 17 Oct the Sqwish organisation on Hugging Face has 0 models, no
sqwishGitHub organisation was found, and the docs treatkettei-coreas an API model. The bases (IFM K2-Horizon-3.7B and 0.9B) are public; the kettei weights are not. Until weights and a licence appear, this is a closed API.- The scores below are for kettei-fat, and the only independent numbers are Benchmark Heaven’s. Everything from Sqwish’s blog is UnverifiedClaim.
kettei (Japanese for “decision”) is the family of decision models Sqwish Labs launched on 14 October 2026 with OneHot, its inference and fine-tuning platform. You send a context (text, images or both), the questions and the allowed answers; a kettei model reads it once and returns a probability for each answer, with no generated text. Sqwish says its earlier work was routers, calibrated judges and context pruners, all “taking a context and a short list of options and returning a probability for each”.
Specs
| Attribute | Value |
|---|---|
| Company | Sqwish Labs (UK). Docs: OneHot is run from the United Kingdom; GPU pools in Europe where possible |
| Launch | 14 Oct 2026 (blog, docs and pricing page dated the same day) |
| Models | kettei-fat (images, fine-tuning “early preview, not available yet”), kettei-chub (images, fine-tuning available), kettei-core (text only, fine-tuning available). Versions “v1.0” for fat and chub |
| Question types | Native /v1/decide: binary, single, ordinal (mix in one request). Jev format via /v1/systemone |
| Endpoints | POST https://api.sqwish.ai/v1/decide (native) and POST https://api.sqwish.ai/v1/systemone (a Jev client “keeps working: change the address and the key”) |
| Input | Up to 65,536 input tokens per request; fat and chub take up to 8 images (JPEG/PNG/WebP, base64) |
| Pricing | Per million input tokens, output not billed: fat $0.06, chub $0.04, core $0.02 (fine-tuned chub $0.08, core $0.04). $5 starter credit, once per email inbox, 30 days |
| Access | New accounts may start waitlisted; most API work needs an active account |
| Tooling | Python SDK sqwishai 0.3.6 on PyPI, agent plugin and MCP tools for Claude Code, Codex and Cursor, a console playground |
| Data | Request content is discarded after the answer and not used to train models offered to others |
| Fallback | fallback: "auto" (default) can let a smaller size answer when a model can’t; the response names the model and any fallback. Set "none" for reproducible comparisons |
Independent boards: Benchmark Heaven (checked 17 Oct 2026)
JevBench, hosted API board (board v1.7.46, updated 16 Oct). Composite is the equal-weight harmonic mean of Intelligence, Calibration, Speed and Cost. Both kettei rows are on the fresh v1.6 regular draw (20261012), full 1,500-item set.
| System | Composite | Intelligence | Calibration | Speed | Cost | $ per 1,000 | p50 |
|---|---|---|---|---|---|---|---|
| kettei-fat (#1) | 77.8 | 80 | 93 | 93 | 58 | $0.026 (est.) | 0.20 s |
| d1-ultra (#2) | 76.1 | 70 | 91 | 89 | 62 | $0.018 (est.) | 0.23 s |
| Sage 1.3.0 (#3) | 74.0 | 66 | 92 | 94 | 58 | $0.025 (est.) | 0.15 s |
| kettei-chub (#4) | 73.5 | 62 | 84 | 95 | 63 | $0.017 (est.) | 0.17 s |
| Jev 1.13.0 (reference, #7) | 71.5 | 64 | 91 | 91 | 55 | $0.032 | 0.24 s |
On the board’s Capability headline (mean of Intelligence and Calibration) fat scores 86.3 (#1, “leads the Jev-class systems”) and chub 73.1. The board counts the 15 of 1,500 items fat refused with HTTP 422 as wrong; the costs use Sqwish’s list prices and the row’s own token counts, output at zero. By subject topic the board’s radar has fat far ahead of Jev 1.13 on rules and policy (88.6 vs 49.4), coding (86.5 vs 64.3), and maths (68.7 vs 10.4), and behind it on “everyday language” (35.0 vs 77.5).
ImageJevBench v0.3.0. kettei-fat is #1 (Capability 75.9, Intelligence 67.4, Calibration 84.5, $0.011 per 1,000); kettei-chub is official #4 (Capability rank 7). For comparison “Sage 1.4 + images” is second by Capability (75.8) and official #5, and “H2O-Lightning-4B (image)” 60.4.
Not on the Decision Index 0.3 board (its private tests are open-model-only, which the Sqwish blog itself notes) or the Artificial Analysis Decision Index. The two boards are never mixed here.
What Sqwish claims (UnverifiedClaim)
- 93.1% on the 231 public JevBench items at 160 ms median from the UK, “more accurate and faster than every other decision model we ran”. The public items are the part of JevBench that can be trained on; the board’s number (77.8 composite) is the independent one.
- Public Decision Index, run by Sqwish with the maintainer’s kit on the stratified 12,000-request sample: fat above every listed decision model, chub matching Perplexity’s 27B Decider, core and dot “leading their size classes”. Not a leaderboard entry.
- Held-out images: 97.5% (fat) and 96.9% (chub) on 5,000 questions Sqwish wrote, against 73.5% for Cloudflare Clef.
- Calibration tests: a hidden-die test (Jev said “1” every time at 83% confidence; kettei stayed near chance) and shuffle and label-swap tests with no flips on answers held at 80% or more. Log-loss figures are Sqwish’s own. No ECE or method for the probabilities is published.
- Games from pixels (ViZDoom, Mario, Tetris) and a “System 1 / System 2” hybrid with Gemini 3.8 Flash: demos by Sqwish, not scored here.
Fit / anti-fit
Fit when you need: a hosted decision model with the strongest independent JevBench result so far, image reading in the same call, a Jev-shaped endpoint for an A/B against Jev, and per-account fine-tuning on chub or core.
Anti-fit when you need: open weights or self-hosting (none found), a GA service with an SLA, a result on Decision Index 0.3 or Artificial Analysis, or fine-tuning on the largest size (fat’s is “not available yet”).
Limits
- Closed weights, base models undisclosed for fat and chub; no published ECE or calibration method.
- Account access can start on a waitlist; the API was not called for this page (no key), so question-type limits beyond the docs were not tested.
- Board rows were measured on a new draw and some at the full 1,500-item set; expect movement as boards re-equate.
- core and dot: no weights, no licence verified.
Why it is in the catalog
Dedicated decision models with a typed API that returns per-option probabilities, public documentation and pricing, and the best independent JevBench and ImageJevBench results at launch. It moves up from the radar’s watch list (it was watch on 15 Oct only because the “open” core and dot weights had not appeared); the page is scored on the API sizes that exist.
Score working (17 Oct 2026, scored on kettei-fat)
- Maturity = availability 6 × 0.30 + docs 8 × 0.25 + integrations 4 × 0.25 + support 3 × 0.20 = 1.80 + 2.00 + 1.00 + 0.60 = 5.40 → 5.4
- Capability = types 9 × 0.30 + calibration 7 × 0.25 + latency 8 × 0.25 + accuracy 8 × 0.20 = 2.70 + 1.75 + 2.00 + 1.60 = 8.05 → 8.1
- Adoption = engagement 20 × 0.35 + ecosystem 8 × 0.35 + production 0 × 0.30 = 7.00 + 2.80 + 0.00 = 9.80 → 10
Availability 6: paid self-serve API, but with an admission step, no GA and no open weights (same as d1-ultra). Docs 8: a full docs site (quickstart, SDKs, interpretation, fine-tuning, residency, llms.txt, OpenAPI). Integrations 4: Jev-compatible /v1/systemone (precedent 3) plus the SDK and agent plugin; no gateway listing. Support 3: company, no SLA.
Types 9: three kinds and images, but I did not run the API or check option caps. Calibration 7: the board’s calibration axis is the top of the API board (93), and Sqwish shows tests, but publishes no ECE (precedent: Sage, Quyet). Latency 8: 0.17–0.20 s measured by the board end to end (independent, no p99). Accuracy 8: one independent board family ranks it first on both JevBench and ImageJevBench (GLiDE / d1-ultra rule); 9 would need a second independent board such as the AA index or DI 0.3.