Lychee-27Bv1.7 · 10 s rule
Data: JevBench v1.7.23, published 2026-10-29 · method v1.7 (10 s rule), measured under the rule
Capability 79.8 · Ranked outside the official caps (no Capability rank)
cost 6.7× Jev (v1.5 reference)
Last measured: 2026-10-13 · measurement release: v1.7.20
CC BY-NC 4.0 (weights; non-commercial use only); base Qwen3.8-27B Apache-2.0
intelligence
74.3
calibration
85.4
speed
83.8
cost
30.0
Composite (secondary)
20.4 · Composite rank #124
Cost and measurement conditions
$0.216 per 1,000 decisions (estimated)
ESTIMATE (*est.), frozen 25 Sep 2026 base-model market reference qwen/qwen3.8-27b: USD 0.42 per million input and USD 3.00 per million output tokens. Measured output tokens: 0 (no generation). Common cost basis: the same 1,485 answered items as other rows on this draw, 763,765 input tokens. USD 0.216014343 per 1,000 decisions (3.34 times the official cap); all-items cost USD 0.544237400 per 1,000 decisions on 1,500 items (1,943,705 input tokens). Estimated model tariff, not charged usage or a host quote; GPU rental is separate.
Measured p50 latency: 0.177 s (on-GPU, no network, loopback; not directly comparable with API latency, which we time end to end from our measuring host, network included)
Ranking and cap use: 0.50 s (est. incl. network (2x + 150 ms))
Self-hosted model: measured p50 0.177 s on-GPU, no network (timed on our GPU over loopback). Not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and latency cap use the same estimate for every self-hosted row: 2 × measured + 150 ms = 0.50 s. Model authors can ask us to time a hosted endpoint instead (on request).
x2 + 0.15 s (assumption, not measured)
evaluator-owned Lium GPU pod (A100 80GB PCIe), offline read-only container (--network none); author's lychee server unchanged at revision f3f95d18 (bf16, one forward pass, linear readout over answer codes, temperature 1.0, prompts read whole; over-length/OOM -> HTTP 422) on the documented stack Python 3.12 / torch 2.14.0 (cu126) / transformers 5.17.0 / flash-linear-attention 0.5.2 / kernels 0.16.1
Measured latency over 1,500 decisions: p50 0.177 s · p95 0.337 s
15 decisions over 10 s (15 of them answered without error, counted as wrong).
A decision that takes longer than 10 seconds (measured time) counts as a wrong answer and as zero calibration for that item: each type's calibration is multiplied by the share of its decisions answered within 10 s. The limit uses the measured time (raw). The displayed adjusted time for self-hosted rows (2 × + 0.15 s) is an assumption and is not used for the limit.