ClassOne Gemma 4 E2B (Gemma-4-E2B backbone with ClassOne decision heads, schema-constrained single forward pass, self-hosted)v1.7 · 10 s rule
Data: JevBench v1.8.18, published 2026-11-20 · method v1.7 (10 s rule), re-scored from the recorded per-decision times
Capability 15.2 · rank #248 of 251 on the open-weights board
Last measured: 2026-10-07 · measurement release: v1.6.1
apache-2.0
intelligence
7.8
calibration
22.6
speed
93.1
cost
85.8
Composite (secondary)
0.5 · Composite rank #236
Cost and measurement conditions
$0.0030 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 gemma, 2B active parameters; throughput 100000 tokens/s (architecture-class calibration, 29 Oct 2026, efficient-engine prefill; capped at 100,000 tokens/s). The row's own scored cost basis: 727,075 input + 0 output tokens over 1,477 decisions = 492.264725796 tokens/decision. GPU-time cost USD 0.00297605281646624/1,000; old list-price cost USD 0.00984529451591063/1,000; applied min(old list-price cost, GPU-time cost) = USD 0.00297605281646624/1,000. 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: 0.029 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 0.029 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)
measured 2026-10-07 on the live v1.6.0/v1.6.1 pool (sha256 901983ae...), 1,477 of 1,500 answered. Weights devops-thiago/classone-gemma4-e2b (sha cd502aa33d288e67f16937c2e125f2297ab7b4ad), Apache-2.0, public and ungated; code github.com/devops-thiago/class-one, Apache-2.0. ONE DOCUMENTED MAPPING OF OURS, AND IT IS THE ONLY ONE. The author's server keys SCORE probabilities '1'..'n' while JevBench's canonical keys are '0'..'n-1'. The UNAMENDED smoke gate REFUSED the first attempt at check score-keys-are-indices, before any sealed item was sent, and the server was torn down with the card verified free. The 1-based convention - which their own schemas.py documents only as 'Probability distribution across each ordered rubric level', without stating the convention - was then established on the AUTHOR'S OWN SERVER with SYNTHETIC requests across every rubric size their own validator allows (2..10): exactly the keys '1'..'n' in order every time, never '0' and never a key above n. It is corroborated independently by their own `score` field being a 1-based probability-weighted mean of the same distribution (1.0185 on a 2-level rubric, 1.6147 on a 10-level one, 3.1155 where the mass sits on key 3; a 0-based convention would have given 0.0185), and a reversed rubric moved the mass to a different key, so 'k' means the k-th level of the criteria array AS SENT - which is what JevBench's 'k-1' means. classone_reindex_r58.ClassOneReindexAdapter therefore relabels score keys k -> k-1 and NOTHING else: the request body, the answer mapping, the noul range check and the choice-must-be-one-of-the-item's-labels rule are the UNCHANGED vendored typesafe adapter's, reused by calling its own run(). An item whose score key set is not exactly {'1'..'n'} is REFUSED and recorded invalid rather than rescored on a guess; none was. choice and noul are untouched. The gate then ran AMENDED in exactly that one check, requiring the key set to be exactly {'1'..'n'} - stricter than the original, because it also pins the count - with the other seven byte-identical, and passed. PREREG-R58-AMENDMENT-1.md (sha256 in PREREG-R58-AMENDMENT-1.md.sha256), receipt PROBE-classone-e2b.json. No file of the author's was edited and no prompt of ours exists. THE SUBMIT FORM CARRIED NO DESCRIPTION FOR EITHER CLASSONE ROW, so everything about how it is run comes from the author's own repository: install line README Quickstart step 1 (`pip install -e ".[dev]"`, in an isolated venv - torch 2.14.1+cu130, transformers 5.19.0, fastapi 0.142.2), serve line step 3 (`uvicorn classone.server.app:app`), and the checkpoint selected the way their own src/classone/server/app.py reads it (CLASSONE_BASE_MODEL). THEIR SERVER REGISTERS POST /v1/systemone ITSELF (app.py:115, next to /v1/decide and /v1/classone) and their DecisionRequest/DecisionResponse in src/classone/schemas.py are the TypeSafe System One shapes, so no transport shim exists anywhere in the path. One RTX 5090 32 GB (Lium, Paris FR). Latency axis uses the standard self-hosted x2 + 0.15 s adjustment. output_tokens is 0 by their own schema comment ('In System 1 models this is 0 because there is no autoregressive decode'). THE 23 LONG ITEMS: on this pool exactly 23 of 1,500 items are 77,857-82,318 tokens and the 24th largest is below 2,048 (run 57's own per-item counts on the identical pool, receipt LONGITEM-TOKENS-R58.json). They are also the 23 items the v1.6.1 cost rule excludes. THIS ROW REFUSES THEM at a limit far below them: all 23 came back as 'HTTP 400: Prompt length (81338 tokens) exceeds maximum limit (8192)', an 8,192-token ceiling the author's code enforces with an explicit error rather than truncating. The 23 invalid are the system's own refusal, counted wrong, and no other item failed. Serial, one decision per request, no retry, no batching, credential-free. add-requests run 58, benchmarkheaven.com/submit ref C4D7AACD (bh-submit 4567a065-e912-40df-9bc1-68fb37cc7bba)
Measured latency over 1,500 decisions: p50 0.029 s · p95 0.043 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.
JevBench v1.6.43 values (method v1.6, before the 10 s rule)
Capability 15.21 · Composite 0.49 · rank #206 on the all-systems board · #163 open weights
intelligence 7.80 · calibration 22.62 · speed 93.29 · cost 70.20
v1.6 speed basis (124 items): p50 0.029 s · p95 0.037 s
These figures are kept for comparison. The current figures above are re-scored under the 10 s rule from the recorded per-decision times and replace them.