Chon-1.0 (12B / E4B / E2B)
Quadrant scores
See the full quadrantScored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.
- Maturity5/10
Apache-2.0 LoRA adapters for Gemma 4 12B, E4B and E2B with an Apache-2.0 server (xoai/chon) that speaks /v1/systemone, plus MPS and CPU modes for the small sizes. Independent author, no hosted API, no SLA.
- Capability8.1/10
JevBench open weights (v1.6.32, 20 Oct): Chon-1.0-12B is #1 of the Jev-class rows, Capability 85.3, Intelligence 79.9 (Jev reference 63.6), Calibration 90.8, p50 0.134 s. Paid submission. Not on Decision Index 0.4 or Artificial Analysis.
- Adoption1/100
About 37 downloads on the 12B repo, no likes, 0 GitHub stars and launch posts on X with under 100 views each (21 Oct). No integrations or production users found.
Vendor claims
- Own measurements: held-out replica (720 items) Intelligence 82.0, calibration proxy 91.4; JevBench public (231) Intelligence 82.8[Vendor claim — not independently verified]
- One H100 80 GB, one request at a time: median 0.11 s, p95 23.5 s; about $0.047 per 1,000 decisions[Vendor claim — not independently verified]
- #1 model on JevBench, surpassing Jev, Clef, Quyet and all other 31B system-one models[Vendor claim — not independently verified]
News covering this model
publicado (OK do Fabricio 21/10 BRT). Scores are calculated from scratch.
What it is
Chon is a family of three LoRA adapters by Ninh Nguyen (@xoai): Chon-1.0-12B on google/gemma-4-12B-it, Chon-1.0-E4B and Chon-1.0-E2B. The base model stays unmodified and is loaded next to the adapter. Given a state and one question with a closed set of options, it returns a probability for every option: choice (2 to 16 options), yes/no, or score. The server speaks the TypeSafe /v1/systemone wire format. English, with Vietnamese also supported. This page is scored on the 12B.
Some decisions take “a short extra pass” that generates text when the model is unsure (the card counts those tokens in output_tokens). That is why the p95 is far above the median. --no-think turns it off.
Board result
| Board | Row | Result |
|---|---|---|
| JevBench open weights (v1.6.32, 20 Oct 2026) | Chon-1.0-12B | #1 Jev-class: Capability 85.3, Intelligence 79.9, Calibration 90.8, $0.050 per 1,000 decisions (1.55× Jev), p50 0.134 s |
| same | Chon-1.0-E4B | Capability 68.4, Intelligence 55.0, Calibration 81.8; p50 22.2 s, outside the Jev-class latency cap |
The board notes a “paid publication exception” for the 12B and lists its base model as undisclosed, while the card names Gemma 4 12B. Not on Decision Index 0.4 or Artificial Analysis (checked 21 Oct); the boards are never combined.
Vendor claims (UnverifiedClaim)
- Own measurements: held-out replica (720 items) Intelligence 82.0, calibration proxy 91.4; JevBench public (231) Intelligence 82.8 (source)
- One H100 80 GB, one request at a time: median 0.11 s, p95 23.5 s; about $0.047 per 1,000 decisions (source)
- “#1 model on JevBench, surpassing Jev, Clef, Quyet and all other 31B system-one models” (post). True for the Jev-class open-weights view on 20 Oct. It is not first on the hosted API board, which is a separate table.
Scores (from scratch, 21 Oct 2026)
- Capability 8.1 = types × 0.30 + calibration × 0.25 + latency × 0.25 + accuracy × 0.20 = 10 × 0.30 + 7 × 0.25 + 7 × 0.25 + 8 × 0.20 = 8.10. Accuracy: 8 × 79.9 ÷ 63.6 = 10.05, capped at 8 (9 needs a second independent board above Jev). Calibration axis 90.8 → 7. Board p50 0.134 s → 7.
- Maturity 5.0 = availability × 0.30 + docs × 0.25 + integrations × 0.25 + support × 0.20 = 7 × 0.30 + 7 × 0.25 + 3 × 0.25 + 2 × 0.20 = 5.00.
- Adoption 1 = engagement × 0.35 + ecosystem × 0.35 + production × 0.30 = 1 × 0.35 + 1 × 0.35 + 0 × 0.30 = 0.70 → 1.