Jeetu Decider 2Bv1.7 · 10 s rule
Data: JevBench v1.7.3, published 2026-10-25 · method v1.7 (10 s rule), measured under the rule
Capability 41.7 · rank #131 of 161 on the open-weights board
Last measured: 2026-10-25 · measurement release: v1.7.2
Apache-2.0 per the model card and LICENSE (LICENSE sha256 prefix 50cbab8a); public, ungated
intelligence
10.2
calibration
73.1
speed
90.9
cost
70.3
Composite (secondary)
1.2 · Composite rank #181
Cost and measurement conditions
$0.0098 per 1,000 decisions (estimated)
Estimated base-model market reference, not charged and not a quote from the author or a host. Frozen pricing module aa2afcd5 maps Qwen/Qwen3.5-2B-Base and Qwen/Qwen3.5-2B to deepinfra:Qwen/Qwen3.5-2B at USD 0.02/M input and USD 0.10/M output before the zero-output convention. Native server usage retained: 726138 input and zero output tokens across 1485 answers, with no generated text. Output is priced at USD 0/M under the native-probability zero-output convention. Server input exceeds the offline count of 721786 by 4352 tokens (+0.60%) across 31 requests; all differences are higher, from 1 to 579 tokens, and occur in requests containing Devanagari or Arabic script. The cause is unverified; no offline substitution or downward correction is applied. Common cost basis excludes the same 15 long items for every token-priced row; refusals still count wrong for capability. Within the official cost cap, with ranked treatment and no cap exception. Cost on the common basis: USD 0.009779636364 per 1000 decisions; official cap USD 0.06459465517241379.
Measured p50 latency: 0.061 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: 0.27 s (est. incl. network (2x + 150 ms))
Self-hosted model: measured p50 0.061 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 = 0.27 s. Model authors can ask us to time a hosted endpoint instead (on request).
x2 + 0.15 s (assumption, not measured)
Author's native server on an evaluator-rented Verda NVIDIA RTX A6000 48 GB, FIN-02 Finland, EU, PAYG non-spot, measured 25 October 2026 about 06:57-07:11 UTC. Pinned local checkpoint and serving code, serial loopback requests, network-isolated read-only container, offline flags and empty provider credentials. Snapshot hashes and threshold verified before held-out input. All 1500 requests completed; node shred and exact-instance teardown verified by the measurement receipts. Approximately USD 0.20 rental is separate from the estimated inference tariff.
Measured latency over 1,500 decisions: p50 0.061 s · p95 0.075 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.
Model and serving disclosure
First measurement of Jeetu456/decider-2b at cd0216309bf590f387500d78d1970af15d04baf5, a Qwen3.5-2B-family fine-tune with a BF16 backbone and FP32 option-pointer decision head. The author's native SystemOne server ran unmodified, with native choice, noul and score probabilities and no generated explanation. The author default abstain threshold 0.55 was frozen before measurement. No answered choice request matched the native English IDK detector, so its choice-only threshold redistribution did not activate. Separately, the scorer records 230 of 373 answered Noul items within its 0.20-0.80 abstention band; this is not attributed to the server's choice threshold. The author documents a threshold-0 alternative, which was not measured. Optional causal-conv1d was omitted in favour of the author's supported F.conv1d fallback, the same maths up to BF16 reassociation; optional-kernel equivalence was not measured. The two weight shards and pointer head share no recorded LFS hashes with 28 compared repositories, including Mapika/decider-2b and Qwen/Qwen3.5-2B. Tokenizer sharing with the base does not imply checkpoint sharing. The model card warns that some When2Call test questions may have been seen during training; training-data contamination against JevBench was not checked. Exactly 1485 answers and 15 preregistered native HTTP 413 over-length refusals; refusals count wrong in the 1500-item denominator, with no evaluator truncation. First 50 responses repeated bit-identically. Earlier provider and installer failures stopped before held-out input and produced no measurement; only the successful recovery attempt is scored.