SD-JEV-4B-v12.1.0 (Kego, author-hosted API)v1.7 · 10 s rule
Data: JevBench v1.8.8, published 2026-11-10 · method v1.7 (10 s rule), measured under the rule
Capability 77.5 · API offering, listed but not ranked on the API leaderboard
Measured on the 27 Oct 2026 draw, which is not yet equated to the reference scale; stronger anchor runs for that draw are being measured. Shown unranked until then.
Last measured: 2026-11-10 · measurement release: v1.8.8
Author-hosted research API (no weights published)
intelligence
66.2
calibration
88.8
speed
81.7
cost
83.6
Composite (secondary)
79.1 · Not ranked on this board
Cost and measurement conditions
$0.0035 per 1,000 decisions (estimated)
GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, EUR 999/month / 730 hours, ECB USD 1.1133/EUR dated 28 Oct 2026, USD 1.5235/hour; average load 70%. Architecture: small qwen, 4B active parameters; throughput 61525 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 533,149 input + 0 output tokens over 1,489 decisions. Generated tokens are counted at the same throughput as prefill. GPU-time cost USD 0.00351839177378149/1,000; old list-price estimate USD 0.0107417528542646/1,000 (deepinfra:Qwen/Qwen3.5-4B at USD 0.03/M input and USD 0.15/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.00351839177378149/1,000. Estimated, not a bill. Customer-stated base Torchcast Decision 4B v1.3 (a Qwen3.5-4B fine-tune) with a LoRA adapter, NOT verified (API-only). Estimate, not a bill and not the author's price; no public tariff exists. Pricing base Qwen/Qwen3.5-4B. The frozen DeepInfra reference was later deprecated; it remains the frozen reference for this estimate. All-items estimate USD 0.00350859401155378/1,000 over all 1,500 items; common-set denominator 1,489. Frozen pricing module sha256 2f52370fe1dba73d75a770a386df350cc76a0898304aeba9e29c91f6e4d3738b.
p50 latency: 0.5 s
none (API)
10 Nov 2026 09:07-09:29 UTC from our server in Germany, network included, author endpoint behind a Cloudflare tunnel. API end-to-end latency: p50 0.544 s, p95 1.242 s, maximum 10.544 s; 1 decision over 10 s. Server-reported usage: 535,592 input / 0 output tokens over 1,495 answered requests. Served model id: SD-JEV-4B-v12.1.0-aligned-s42-step0256.
Measured latency over 1,500 decisions: p50 0.544 s · p95 1.242 s
1 decision over 10 s (0 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.
Model and serving disclosure
Measured through the author's own hosted endpoint (one NVIDIA A10, author-operated). The author states the base is Torchcast Decision 4B v1.3 with a LoRA adapter; we could not verify the weights. The author used the 231 public JevBench items during development, which the rules allow; the sealed draw was not available to them. The author's endpoint logs request bodies, so the author's systems received the items of this run. Speed is end-to-end latency to the author's endpoint, including network. Requests over the endpoint's 16,384-token input limit, which the author raised from 4,096 before the run, were refused and are counted as wrong answers. Calibration is computed on the 1,495 answered items, as for every row with refused requests. Exactly 5 of 1,500 items (about 76,000-80,000 tokens each) were refused; this count was registered before the run.