Vega 4B
Data: JevBench v1.6.40, published 2026-10-22
Capability 37.7 · rank #117 of 139 on the open-weights board
Last measured: 2026-10-10 · measurement release: v1.6.5
vegaml package Apache-2.0; HF checkpoint repo declares no licence field (implied, not declared); backbone Qwen3.5 Apache-2.0
intelligence
16.7
calibration
58.7
speed
84.4
cost
47.2
Composite (secondary)
3.6 · Composite rank #147
Cost and measurement conditions
$0.058 per 1,000 decisions (estimated)
v1.5-M2 base-model reference: Vega 4B is a decision engine on a frozen Qwen3.5-4B backbone (Qwen/Qwen3.5-4B, resolves in pricing_v15 to deepinfra:Qwen/Qwen3.5-4B). DeepInfra list price USD 0.03/M input, USD 0.15/M output from the dated DeepInfra catalog snapshot 2026-09-25 (jevbench-v15-measure-20260925/receipts/DEEPINFRA-MODELS-20260925.json, sha256 84eb7926acd4e93b71d695be24898e2ca999b20011d1b66cc0348ce7c6c3546f; source https://api.deepinfra.com/models/list), read by the frozen pricing_v15 on 2026-10-10. The listing is flagged by pricing_v15 as deprecated after the 25 Sep cut-off (DEPRECATED_AFTER_CUTOFF) and stays frozen with that note. The model generates no text: 0 output tokens, cost = input tokens x USD 0.03/M.
Measured p50 latency: 0.190 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: 0.53 s (est. incl. network (2x + 150 ms))
Self-hosted model: measured p50 0.190 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 = 0.53 s. Model authors can ask us to time a hosted endpoint instead (on request).
x2 + 0.15 s (assumption, not measured)
evaluator-owned Lium RTX A6000 48 GB pod (Des Moines, US), offline after setup (egress cut and proven), reviewed vegaml==0.6.1 wheel, backbone bf16 / engine fp32, context 73,728 (longer states cut on the right), reference PyTorch path for linear attention (no flash-linear-attention / causal-conv1d)