ENPT

Quyet 1.0

Last updated:

Open180–380ms$0/M input

Quadrant scores

See the full quadrant

Scored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.

  • Maturity5.4/10

    Five Apache-2.0 checkpoints (31B to 153M) with a pip package (quyet 1.0.1) that builds the prompt and applies the fitted temperatures; Large also loads in Transformers and vLLM. Answers in the /v1/systemone shape. Solo author, cards only, no SLA.

  • Capability8.1/10

    Choice / Score / Noul, at most 10 options per question. Large: #1 on the JevBench v1.6.1 open-weights board (composite 71.4, capability 81.7, Calibration 90.0, raw p50 0.38 s) and rank 6 group on Decision Index 0.3 (58.71). Encoder sizes score near zero there.

  • Adoption5/100

    Three days old: Large ~240 downloads / 20 likes; community GGUFs (Medium ~6.8k, Large ~1.2k downloads). No launch post, integrations or production users found.

Vendor claims

Read this first

  1. One family, very different members. The two decoders (Large, Medium) read option-letter logits. The three encoders (Small, Small-EN, Tiny) use fixed heads and score near zero on JevBench’s Intelligence axis. The scores on this page are for Large.
  2. The author publishes no accuracy numbers. Everything below comes from Benchmark Heaven and the Decision Index, not from the cards.
  3. At most 10 options per question, and Large truncates the state at about 6,000 tokens inside an 8,000-token prompt.

The family

Model Base Params Readout JevBench v1.6.1 (open board) Decision Index 0.3
Quyet-1.0-Large Gemma-4-31B-it, merged LoRA r16 31.3B option letters 71.4 (#1), capability 81.7 (#1) 58.71 (rank 6 group)
Quyet-1.0-Medium Qwen3.5-4B 4.7B option letters 43.6 (#15), capability 59.2 —
Quyet-1.0-Small SEA-LION-ModernBERT-300M 328M fixed heads 0.1 (#90) —
Quyet-1.0-Small-EN ModernBERT-base 153M fixed heads 0.3 (#82) —
Quyet-1.0-Tiny mmBERT-small (16 layers), distilled from Medium 183M option-query head 0.1 (#91) —

All five went up on Hugging Face on under Apache-2.0. pip install quyet, then quyet.load("chinhnc/Quyet-1.0-Large").

Large on JevBench v1.6.1 (verified 7 Oct 2026)

Composite 71.4 (95% interval 69.8–72.2), Intelligence 73.4, Calibration 90.0, Speed 86.9, Cost 50.5. $0.045 per 1,000 decisions (1.38× Jev, inside the 2× cap), raw p50 0.38 s. Jev 1.13.0 is the unranked reference at 71.5 / capability 77.1. Medium: Intelligence 41.7, Calibration 76.8, p50 0.30 s. The encoders fall under the board’s Intelligence gate (Intelligence 3.7–6.1), so their composites are near zero.

Fit / anti-fit

Fit when you need: a top open decision model on one 80 GB GPU (Large), Apache-2.0 weights, Vietnamese as well as English, or a CPU-sized encoder for narrow, cheap routing (with low accuracy on general decisions).

Anti-fit when you need: more than 10 options, long states, published accuracy or ECE from the author, or a supported product.

Why it is in the catalog

Decision-trained open weights with a typed interface and calibrated temperatures, and the Large size leads an independent board.

Score working (7 Oct 2026, scored on Large)

  • Maturity = availability 8 × 0.30 + docs 6 × 0.25 + integrations 5 × 0.25 + support 1 × 0.20 = 2.40 + 1.50 + 1.25 + 0.20 = 5.35 → 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 6 × 0.35 + ecosystem 8 × 0.35 + production 0 × 0.30 = 2.10 + 2.80 + 0.00 = 4.90 → 5

Types is 9, not 10, because of the 10-option cap. Calibration stays at 7: the board’s calibration axis is strong, but the author publishes no ECE.