torchcast-decision-12b (Torchcast AI, Gemma-4-12B fine-tune, option-letter logprob readout)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 71.7 · rank #67 of 251 on the open-weights board
Last measured: 2026-10-04 · measurement release: v1.6.0
cc-by-nc-4.0
intelligence
60.5
calibration
82.9
speed
91.3
cost
56.4
Composite (secondary)
69.8 · Composite rank #85
Cost and measurement conditions
$0.028 per 1,000 decisions (estimated)
v1.5.8 fast-lane delivery row (combined v1.5.8 preview, not yet published) cost per 1,000 decisions carried (pricing rules unchanged; v1.6 item lengths differ): ESTIMATE (v1.5-M2 base-model reference; no public tariff, customer asked for the base-model reference): OpenRouter google/gemma-3-12b-it list price $0.05/M input, $0.15/M output (gemma-4-12b-it is not listed on OpenRouter, read 2026-10-04; same reference as the Cygnet and Jev-Omni Gemma-4-12B rows) x the shim's own measured usage.prompt_tokens/completion_tokens; self-hosted on 1x H100 80GB (Lium), vLLM 0.30.0 BF16; estimated, not charged
Measured p50 latency: 0.045 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: 0.24 s (est. incl. network (2x + 150 ms))
Self-hosted model: measured p50 0.045 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 = 0.24 s. Model authors can ask us to time a hosted endpoint instead (on request).
x2 + 0.15 s (assumption, not measured)
evaluator-owned Lium GPU pod, 1x NVIDIA H100 80GB, weights fetched in a separate model-only container, scored run inside docker --network none (loopback only, no secrets, label-free inputs); pod destroyed afterwards; paid public fast-lane evaluation, 4 Oct 2026 (paid evaluation)
Measured latency over 1,500 decisions: p50 0.045 s · p95 0.078 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.
Breakdowns (not part of the score)
Chance-corrected competence (0 = chance, 100 = perfect) by category on S+P+L1+L2+L3. Raw values, not equated; the same cells feed the compare view.
Capability by subject topic
| Rules, policy & law | 57.5 | n=1616 |
|---|---|---|
| Coding & software | 72.4 | n=382 |
| Math & numbers | 22.4 | n=370 |
| Finance & commerce | 59.1 | n=223 |
| Support & operations | 61.0 | n=160 |
| Everyday language | 76.3 | n=145 |
| Safety & security | 77.1 | n=109 |
Use cases (TypeSafe categories)
| Model routing | 74.2 | n=524 |
|---|---|---|
| Legal & compliance | 64.0 | n=515 |
| Customer support | 56.8 | n=240 |
| E-commerce | 56.3 | n=119 |
| Insurance claims | 61.5 | n=119 |
| Risk assessment | 53.3 | n=118 |
| Financial crime | 41.8 | n=101 |
| Feature extraction | 35.7 | n=99 |
| Lead generation | 68.7 | n=95 |
| Recruiting | 30.5 | n=94 |
| Knowledge graphs | 52.6 | n=94 |
| LLM guardrails | 80.8 | n=93 |
| Moderation | 60.5 | n=93 |
| Code linting | 61.5 | n=90 |
| Search & retrieval | 85.5 | n=90 |
| Science | 33.7 | n=90 |
| Advertising | 49.5 | n=88 |
| Gaming | 20.3 | n=86 |
| Demand forecasting | 0.0 | n=83 |
Cells need at least 30 answered items; † fewer than 30 (indicative only; drawn as hollow, lighter points on the compare radars).
JevBench v1.6.43 values (method v1.6, before the 10 s rule)
Capability 71.74 · Composite 69.91 · rank #27 on the all-systems board · #11 open weights
intelligence 60.54 · calibration 82.93 · speed 91.75 · cost 56.36
v1.6 speed basis (124 items): p50 0.037 s · p95 0.074 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.