Starlord-27B-v2.0-previewv1.7 · 10 s rule

Data: JevBench v1.8.12, published 2026-11-16 · method v1.7 (10 s rule), measured under the rule

Capability 84.5 · Not ranked in the current release

Added after the v1.8.10 ranking review; it is ranked at the next review.

Last measured: 2026-11-15 · measurement release: v1.8.12

cc-by-nc-nd-4.0 (model card); base Qwen/Qwen3.8-27B

Base model: Qwen/Qwen3.8-27Bmodel card

intelligence

75.4

calibration

93.5

speed

88.1

cost

61.1

Composite (secondary)

77.4 · Not ranked on this board

Cost and measurement conditions

$0.020 per 1,000 decisions (estimated)

GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, USD 1.5235/hour; average load 70%. Architecture: dense qwen, 27B active parameters; throughput 11763 tokens/s (architecture-class calibration, efficient-engine prefill). GPU-time cost USD 0.019849097845594/1,000; old list-price estimate USD 0.162205748824715/1,000 (qwen/qwen3.8-27b at USD 0.42/M input and USD 3.0/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.019849097845594/1,000. Estimated, not a bill.

Measured p50 latency: 0.085 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: 0.32 s (est. incl. network (2x + 150 ms))

Self-hosted model: measured p50 0.085 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 = 0.32 s. Model authors can ask us to time a hosted endpoint instead (on request).

x2 + 0.15 s (assumption, not measured)

Evaluator-owned Verda GPU VM (RTX PRO 6000), offline read-only container (--network none); author's inference/serve_systemone.py unchanged at model commit eb4b2c81d89d32b00bfdeaf2501dcff46afe8406. One BF16 model instance, candidate-logit readout (FP32 softmax, T=1.0), 0 generated tokens, no KV cache. Full-attention layers run the author's two-block bidirectional FlashAttention (flash-attn 2.8.3 varlen: request prefix attends to the prefix, the rest to the whole request); linear-attention (GDN) layers are causal via fla 0.4.1 kernels. Limit 131,072 tokens per joint request, no truncation, HTTP 422 above the limit or on GPU OOM. PyTorch 2.10.0 cu130, transformers 5.3.0, flash-attn 2.8.3, Python 3.12 and decision_index base classes at commit 25c6ec7b.

Measured latency over all decisions: p50 0.085 s · p95 0.165 s

— decisions over 10 s.

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

Paid public evaluation. Measured on the 27 Oct draw on an evaluator-owned rented GPU with the author's serving code unchanged at the pinned model commit, offline (no network). Candidate-logit readout with 0 generated tokens. Failed or unanswered requests count as wrong answers. The model card says the public JevBench split was used for evaluation and tuning, not training.

Published source