StartLux-Decision v1.1 27B AutoThink (StartLux Labs, Qwen3.8-27B fine-tune, gate 0.6)
Data: JevBench v1.6.39, published 2026-10-22
Capability 80.3 · Not ranked in the current release
cost 11.0× Jev (v1.5 reference)
Last measured: 2026-10-16 · measurement release: v1.6.25
Weights CC BY-NC 4.0 (non-commercial: commercial use needs a separate licence from StartLux Labs; listing a measured score is evaluation use); inference package Apache-2.0
intelligence
69.0
calibration
91.7
speed
75.1
cost
23.5
Composite (secondary)
10.9 · Composite rank #128
Cost and measurement conditions
$0.354 per 1,000 decisions (estimated)
ESTIMATE (*est.), base-model market reference through the frozen pricing_v15 module: base Qwen/Qwen3.8-27B -> qwen/qwen3.8-27b USD 0.42/M input, USD 3.0/M output. Estimated, not charged, not a quote from the author or any host. Frozen pricing_v15: base-model references unchanged since sha256 3948e3f1...ea4d0, the module this row was first priced with (the later re-freeze only adds a hosted-API launch-list key), USD per 1,000 decisions over the common 1,485-item cost basis (the 15 very long items are left out of cost for every row; refused items cost nothing and still count as wrong answers). The model cards do not name the Qwen version; Qwen/Qwen3.8-27B is the base the Decision Index metadata gives for the sibling Instant variant (AutoThink itself is not on the Decision Index), shares the same NOTICE, consistent with the NOTICE (modified Alibaba Cloud Apache-2.0 model) and with the config (Qwen3_5ForConditionalGeneration, 64 layers, hidden 5120), the same reference as the live torchcast-decision-27b row (a fine-tune of StartLux-Decision-27B v1). OVER THE OFFICIAL COST CAP, DISCLOSED: USD 0.354306 per 1,000 decisions on the common basis (USD 0.350763 over all 1,500 items) is 5.485x the official cap USD 0.064595; the row is ranked (v1.6.16 vllm-sr-d3 and v1.6.24 imajev-9b precedent) and falls outside the official-caps board views by construction. COMMON COST BASIS (v1.6.30): cost per 1,000 decisions is computed on the common 1,485 items of this draw, leaving out the 15 very long items (at least 76,358 input tokens as counted by the rows that answered them). They are the 15 items the Jev reference refused with max_tokens_exceeded, which the other rows that did not answer them refused or failed on; their tokens count for no row, whether the row refused or answered them, and a refused item still counts as a wrong answer for Intelligence. This row: USD 0.354306 per 1,000 decisions (5.485x the official cost cap USD 0.064595), against USD 0.350763 over all 1,500 items (v1.6.29).
Measured p50 latency: 0.056 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: 0.26 s (est. incl. network (2x + 150 ms))
Self-hosted model: measured p50 0.056 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 = 0.26 s. Model authors can ask us to time a hosted endpoint instead (on request).
x2 + 0.15 s (assumption, not measured)
Evaluator-owned rented Lium GPU pod (measurement receipts on file), 1x NVIDIA H200 141 GB (sm_90, driver 580.173.02), Finland (EU). The author's own server on vLLM 0.19.1, offline, loopback, one request at a time.