Philly Decider v2 (PhillyLabs)v1.7 · 10 s rule
Data: JevBench v1.8.15, published 2026-11-17 · method v1.7 (10 s rule), measured under the rule
Capability 69.3 · rank #84 of 243 on the open-weights board
Last measured: 2026-10-26 · measurement release: v1.7.10
Apache-2.0 per the model card and the repository's LICENSE file (H2O.ai's licence file, kept unchanged); base h2oai/h2o-lightning-4b Apache-2.0; backbone Qwen/Qwen3.5-4B Apache-2.0; ungated
intelligence
56.1
calibration
82.5
speed
93.2
cost
79.1
Composite (secondary)
75.0 · Composite rank #19
Cost and measurement conditions
$0.0050 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: 752,008 input + 0 output tokens over 1,485 decisions = 506.402693603 tokens/decision. GPU-time cost USD 0.00497606795343036/1,000; old list-price cost USD 0.0151920808080808/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00497606795343036/1,000. The GPU-time estimate is lower. Estimated, not a bill. Intelligence, Calibration, measured latency and the common cost-item membership are unchanged. Prior prices remain in the archived v1.7 releases.
Measured p50 latency: 0.024 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: 0.20 s (est. incl. network (2x + 150 ms))
Self-hosted model: measured p50 0.024 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 = 0.20 s. Model authors can ask us to time a hosted endpoint instead (on request).
x2 + 0.15 s (assumption, not measured)
Evaluator-owned rented Verda GPU VM, one NVIDIA RTX PRO 6000 Blackwell Server Edition 96 GB (97,887 MiB, driver 580.178.04), FIN-03 (Finland, EU), PAYG non-spot, instance type 1RTXPRO6000.30V, image 24.04.cuda13.0, USD 2.583095890410959/h recorded by the runner, 150 GB disk. One VM for this row, one measured attempt. Measured 26 Oct 2026, VM window about 19:32-19:43Z (639.6 s from create to closed), about USD 0.46 GPU rental (USD 0.459 at the recorded hourly rate). Author server (stock vLLM at 127.0.0.1:8000 behind the author's shim at 127.0.0.1:8741) in a --network none --read-only container, network cut before sealed input (cut proof and offline model-hash proof receipts: only loopback, no default route, DNS and TCP failed; all 17 model files verified), read-only model, source and input, private output, HF offline flags, no credentials. Before sealed input: the image digest, the stack versions, the Triton gate, the weights fetch with a node re-verify (17 of 17), the pre-sealed smoke in its own server process (/health equal to the preregistered configuration, a cold trimmed probe at the longest original state, anchors from 83 to 40,959 tokens, over-limit and 256-option refusals with recovery, typed probes) and the TTL fit gate all passed (forecast of items over 10 s: 0; forecast compute 674 s); a public pre-warm preceded the scored loop, which ran in a fresh server process. Input sha256 dfd36720077ef7c41ace367524896507c7f44ce71306fb3e659e8b85069d995b (v1.6-regular-20261012) was re-checked inside the container (match). Complete 1,500-row run, rc 0; scored loop 59 s. Raw output sha256 6d02e677e68cd68383a789e4f720e025b4c14845a02c831a2ba95f8b6f7eaf67 (the pull-hash receipt). Node-shred and local-shred receipts verified; exact-id teardown receipt confirms the VM closed. GPU rental is separate from the benchmark's estimated model tariff.
Measured latency over 1,500 decisions: p50 0.024 s · p95 0.046 s
0 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.