ENPT

Lychee-27B

Last updated:

Open—$0/M input

Quadrant scores

See the full quadrant

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

  • Maturity4/10

    Weights are public but under CC BY-NC 4.0 (non-commercial), derived from Apache-2.0 Qwen3.8-27B with a NOTICE. A short card with a server (POST /v1/systemone), a Python class and a 262,144-token context; one 80 GB GPU. No hosted API, one anonymous author, no SLA.

  • Capability6.6/10

    Choice, Noul, Score from one forward pass. Full 62.99 on the official Decision Index 0.3.2 (#8; Jev 60.11). No calibration method or ECE and no latency number on the card, so those axes stay low. The card's public 62.8 matches the board's 62.76.

  • Adoption1/100

    Five days old: 43 downloads, 0 likes, one HF repo, no GGUF, no post, no GitHub repo and no organisation page found. No integrations or reported users.

Vendor claims

  • Decision Index 0.3 public index 62.8 against Jev 58.0, measured with the board's kit against the server in the repository; Knowledge & Reasoning 51.8, Language 66.2, Retrieval 63.9, Tools 80.8, Arts 47.0[Vendor claim — not independently verified]

publicado (OK do Fabricio 17/10 BRT).

Read this first. The weights are not open for commercial use: the license is CC BY-NC 4.0. It is public and ungated, and the Apache-2.0 base (Qwen3.8-27B) carries a LICENSE.apache-2.0 and a NOTICE that lists the changes. We list it because it meets the catalog’s public-availability test and has an independent top-10 Full score, and we mark the license limit as the first anti-fit.

Lychee-27B is a decision model fine-tuned from Qwen3.8-27B by LycheeAI (one repository, no organisation page). A 255-way linear readout over single-token answer codes replaces the generation head, and the full-attention layers read the whole input (only padding is masked), so nothing is generated. You send a state and typed questions (choice, noul, score) and get a probability for every option.

Specs

Attribute Value
Developer LycheeAI (Hugging Face user; no GitHub or site found)
Base Qwen3.8-27B (Apache-2.0), fine-tuned; readout in readout.safetensors
License CC BY-NC 4.0 for the modifications and code (non-commercial); base under Apache-2.0
Weights ~52 GB bf16; one 80 GB CUDA GPU
Types choice, noul, score (levels as a list, lowest first)
Interface python -m lychee.server serves POST /v1/systemone with {state, questions}; or from lychee import Lychee
Context 262,144 tokens; longer prompts, or ones that do not fit in memory, are refused, never truncated
Released 12 Oct 2026 (card updated 13 Oct)

Official board: Decision Index 0.3.2

Data generated 16 Oct 2026 19:00 UTC. Full = 0.2 public + 0.5 same-skills + 0.3 new-domains.

Lychee-27B Jev 1.13
Full 62.99 (#8) 60.11 (#18)
Public 62.76 (#13) 57.96 (#37)

The card’s public 62.8 is the author’s run with the board’s kit and matches the board. Not on JevBench or Artificial Analysis as of 17 Oct (we found no row).

What is vendor-run

  • The area scores on the card (Language 66.2, Retrieval 63.9, Tools 80.8, Arts 47.0, Knowledge 51.8) are the author’s. The board’s Full score is the independent number.
  • The card gives no calibration method, ECE or latency; nothing here is a claim we could check.

Fit / anti-fit

Fit: research and internal evaluation of a top-10 decision model with a Jev-shaped endpoint, on one 80 GB GPU.

Anti-fit: any commercial use (CC BY-NC 4.0), a hosted API, support, published calibration or latency numbers.

Limits

  • Non-commercial license; one anonymous author; no code outside the HF repo.
  • No latency, calibration or training-data disclosure on the card.
  • A short card: no disclosure of training data or decontamination.

Why it is in the catalog

Public weights with a typed Choice / Noul / Score interface in the System One shape and an official-board Full score above Jev’s. The license limit lowers availability but does not fail the inclusion test (publicly available, minimal API documentation).

Score working (17 Oct 2026)

  • Maturity = availability 6 × 0.30 + docs 5 × 0.25 + integrations 3 × 0.25 + support 1 × 0.20 = 1.80 + 1.25 + 0.75 + 0.20 = 4.00 → 4.0
  • Capability = types 10 × 0.30 + calibration 4 × 0.25 + latency 4 × 0.25 + accuracy 8 × 0.20 = 3.00 + 1.00 + 1.00 + 1.60 = 6.60 → 6.6
  • Adoption = engagement 1 × 0.35 + ecosystem 1 × 0.35 + production 0 × 0.30 = 0.35 + 0.35 + 0.00 = 0.70 → 1

Availability 6: public ungated weights, but non-commercial (precedent: StartLux, bev-decider at 7; lower here because the card is short and the server is the only path). Docs 5: a short card with a quickstart and the license notice. Integrations 3: /v1/systemone (precedent). Support 1: one anonymous author.

Calibration 4 and latency 4: no method, ECE or latency number (the “no number” precedent). Accuracy 8: independent board result (Full 62.99, #8).

JevBench row and Decision Index 0.4.1 (30 Oct 2026)

publicado (OK do Fabricio 30/10 BRT).

JevBench v1.7.20 (29 Oct) published Lychee-27B, measured on 13 Oct as a paid priority evaluation at the author’s request. It is ranked.

Board Row Result
JevBench (Benchmark Heaven, v1.7.20) Lychee-27B Capability Score 79.8 (#21), Intelligence 74.3, Calibration 85.4, Cost 57.3 ($0.026 per 1,000 decisions, 0.82× Jev); measured p50 0.177 s on GPU
Decision Index 0.4.1 (29 Oct) Lychee-27B Full 67.35 (#18 of 247), public 62.76; ECE 0.108; median 97 ms

Scores unchanged (4.0 / 6.6 / 1). These rows are recorded as facts; the score was reviewed on 30 Oct and kept at 6.6.