shisa-de-1 (shisa-ai, Gemma-4-26B-A4B)
Data: JevBench v1.6.40, published 2026-10-22
Capability 65.5 · rank #35 of 139 on the open-weights board
Last measured: 2026-10-16 · measurement release: v1.6.26
Apache-2.0
intelligence
49.9
calibration
81.2
speed
92.8
cost
49.8
Composite (secondary)
62.4 · Composite rank #31
Cost and measurement conditions
$0.047 per 1,000 decisions (estimated)
ESTIMATE (*est.), base-model market reference through the frozen pricing_v15 module: declared base google/gemma-4-26b-a4b-it -> google/gemma-4-26b-a4b-it USD 0.09/M input, USD 0.3/M output (frozen at the 25 Sep cut-off, I-3). Estimated, not charged, not a quote from the author or any host. Frozen pricing_v15: base-model references unchanged since sha256 dcfcecb0...2835 (the later re-freezes only add or re-price hosted-API launch-list keys; this row was first priced with 3948e3f1...d0), 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 or failed items cost nothing and still count as wrong answers). Tokens as reported by the evaluator's adapter (prompt tokens of the request(s); output_tokens 0 - the single sampled token is discarded, the answer is the logprob readout). Base string is the frozen module's lower-case key (the card spelling raises). 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 0.047250 per 1,000 decisions (0.731x the official cost cap USD 0.064595), against USD 0.046777 over all 1,500 items (v1.6.29).
Measured p50 latency: 0.031 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 0.031 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)
Evaluator-owned rented Lium GPU pod (measurement receipts on file), 1x NVIDIA H100 80 GB HBM3 (sm_90, driver 580.178.04, United States node, pre-registration Amendments 1-2) - the card most of these systems were first evaluated on, but a different pod, stack and pin set from the earlier runs, so Speed is as measured on this card and probabilities may differ from the earlier runs in the last digits. Python packaging: uv venv from a hashed lock at the old pins, dependencies resolved as of the old build time where the old freeze is lost (declared deviation, UNVERIFIED where noted). Isolation: author server started inside a network namespace with loopback only, the cut proven before and after the sealed input was staged, synthetic smoke gate passed in between; runner unchanged (run_v15.py --resume --resume-keep-failures); one use of input-selfhosted-1500 (custody #20324, under the pre-registration). Published Speed uses the standing x2 + 0.15 s self-hosted adjustment. Stock vLLM server read through the evaluator's shisa_vllm.py adapter.