Torchcast Decision 4B (v1.3)
Data: JevBench v1.6.42, published 2026-10-23
Capability 77.5 · rank #11 of 140 on the open-weights board
Last measured: 2026-10-23 · measurement release: v1.6.42
Apache-2.0 (weights, from v1.2); base Qwen/Qwen3.5-4B Apache-2.0; serving package Apache-2.0
intelligence
67.1
calibration
87.8
speed
92.6
cost
64.8
Composite (secondary)
76.2 · Composite rank #1
Cost and measurement conditions
$0.015 per 1,000 decisions (estimated)
Self-hosted open weights, no bookable tariff. v1.5-M2 base-model reference rule: base Qwen/Qwen3.5-4B (model card base_model, finetune). OpenRouter snapshot of 2026-10-23 lists no Qwen3.5-4B; binding precedent INTERPRETATION-1 I-2/I-3 uses DeepInfra Qwen/Qwen3.5-4B at USD 0.03/M input and USD 0.15/M output from the frozen 25 Sep 2026 price cut-off snapshot (read 2026-10-23), through the frozen pricing_v15 module (sha256 aa2afcd5a58d0c4cc0e3d6867d590c70b6ac58e971410c27faf9628f97bcf8b7), the same reference as the live v1.2 row. ESTIMATE, not charged, not a quote from the author or any host. No-generation readout: output tokens 0. Frozen reference deepinfra:Qwen/Qwen3.5-4B; 982,889 input / 0 output tokens over all 1,500 decisions (USD 0.0196578 per 1,000 over all 1,500 items; the 15 truncated prompts count at their truncated size of about 16,370 tokens each), inside the official cap. 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 that count for no row; 737,326 input / 0 output tokens on those 1,485 items. This row: USD 0.014895 per 1,000 decisions (0.231x the official cost cap USD 0.064595), against USD 0.019658 over all 1,500 items.
Measured p50 latency: 0.033 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: 0.22 s (est. incl. network (2x + 150 ms))
Self-hosted model: measured p50 0.033 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 = 0.22 s. Model authors can ask us to time a hosted endpoint instead (on request).
x2 + 0.15 s (assumption, not measured)
Evaluator-owned rented Verda GPU VM (1A100.22V, FIN-02, Finland, EU, pay-as-you-go), 1x NVIDIA A100 SXM4 80GB; offline read-only container (network none); author's torchcast_decision server unchanged at tag v1.3.0 with the bundled decision_config.json (bf16, one forward pass reading option letters, non-causal full attention as trained, calibration temperature 0.9, yes/no commit floor 0.801) on the documented stack torch 2.9.1+cu128 / transformers 5.8.1 / flash-linear-attention 0.5.2 / causal-conv1d 1.7.0 (fast kernels active; CUDA graphs with a key-padding mask in non-causal mode). Driven by the unchanged self-hosted runner on the frozen v1.6-regular-20261012 self-hosted input (sha256 dfd36720...), measured 23 Oct 2026; raw latency p50 0.034 s, p95 0.056 s, maximum 7.73 s. The standard self-hosted x2 + 0.15 s adjustment applies (assumption, not measured).