ClassOne Qwen 3.5 2B (Qwen3.5-2B backbone with ClassOne decision heads, schema-constrained single forward pass, self-hosted)
Data: JevBench v1.6.39, published 2026-10-22
Capability 19.2 · rank #133 of 139 on the open-weights board
Last measured: 2026-10-10 · measurement release: v1.6.9
apache-2.0
intelligence
5.8
calibration
32.6
speed
93.1
cost
70.2
Composite (secondary)
0.2 · Composite rank #167
Cost and measurement conditions
$0.0098 per 1,000 decisions (estimated)
ESTIMATE (*est.), exact base-model market reference; self-hosted open weights with no per-token tariff of their own. The unmodified frozen pricing_v15 resolves the author's own declared base spelling Qwen/Qwen3.5-2B straight through to deepinfra:Qwen/Qwen3.5-2B at USD 0.02 per 1M input and USD 0.1 per 1M output in the dated 25 Sep 2026 snapshot. No new release-lane pricing line. OUTPUT TOKENS ARE 0 BY CONSTRUCTION, not by estimate: the readout is a distribution over the item's own candidate labels from one forward pass and nothing is generated, so the row's cost rests on the input rate alone. This row's own usage on draw v1.6-regular-20261010-a2: 735,765 input and 0 output tokens over 1,500 decisions. Estimated, not charged.
Measured p50 latency: 0.030 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: 0.21 s (est. incl. network (2x + 150 ms))
Self-hosted model: measured p50 0.030 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 = 0.21 s. Model authors can ask us to time a hosted endpoint instead (on request).
x2 + 0.15 s (assumption, not measured)
The author's own README Quickstart install line (pip install -e '.[dev]') at code pin github.com/devops-thiago/class-one@6cca132b and their own serve line (uvicorn classone.server.app:app); their server registers POST /v1/systemone itself, so the request and answer shapes are theirs. CLASSONE_BASE_MODEL was pointed at the LOCAL PINNED SNAPSHOT DIRECTORY, which their own code accepts, so the revision cannot drift to main and classone_heads.pt resolves locally with no request-time hub fetch. Their hard-coded MAX_TOKENS_PER_REQUEST = 8192 refuses longer prompts with HTTP 400; those items are recorded as invalid, the same way for every system. ONE documented mapping of ours and the only one: their server keys score probabilities '1'..'n' while JevBench's canonical keys are '0'..'n-1', so run 58's byte-identical ClassOneReindexAdapter relabels score keys k -> k-1 by calling the unchanged vendored typesafe adapter's own run(), and an item whose key set is not exactly {'1'..'n'} is REFUSED rather than rescored on a guess; choice and noul are untouched. The two already-published ClassOne rows carry exactly this treatment. evaluator-owned Lium RTX 5090 32 GB pod (pod calm-wolf-31, USD 0.72/h; Brasov, Romania, EU by the RIPE whois of its public IP, receipt r90/receipts/POD-IP-WHOIS-R90.txt); pod-level iptables egress cut installed, verified and TIMESTAMPED pod-side (logs/EGRESS-CUT-R90.log) BEFORE the sealed input was uploaded, so the ordering is provable from pod evidence and not only from the session transcript; HF/transformers offline, no credentials in the environment, one decision per request, no batching, no retry. The five rows ran strictly one at a time on the one card, each server started with setsid and stopped by its exact process group with a card-free check in between.