Liquid AI d1-3B
Data: JevBench v1.6.40, published 2026-10-22
Capability 49.5 · rank #90 of 139 on the open-weights board
Last measured: 2026-10-10 · measurement release: v1.6.5
LFM Open License v1.0 (commercial use only below USD 10 M annual revenue)
intelligence
22.1
calibration
76.9
speed
94.3
cost
50.6
Composite (secondary)
8.9 · Composite rank #135
Cost and measurement conditions
$0.044 per 1,000 decisions (estimated)
ESTIMATE (*est.), size-class reference; self-hosted open weights, no bookable listing for this base. The declared base LiquidAI/LFM2.5-VL-3B is not a key of any frozen pricing table, and the frozen 25 Sep 2026 snapshot's only Liquid entry is liquid/lfm-2.5-2.6b:free at (0, 0) - a free route, which may not price a row (Florian, free tiers: use the published paid price or the estimate). Method, as aplomb-1, wald-4b-v2, jiwo-* and the ClassOne trio carry: the DeepInfra Qwen3.5 size-class reference from the frozen snapshot, rounded ONE CLASS UP so the estimate can only OVERSTATE this row's cost - deepinfra:Qwen/Qwen3.5-4B at USD 0.03 per 1M input and USD 0.15 per 1M output, for a 3B model. Applied to this row's own measured usage: 2,222,971 input tokens and 0 output tokens over 1,500 decisions. The 2B rate (0.02, 0.10) would give 0.029639613333333335 and was deliberately NOT used because it understates. Precedent for a LiquidAI base priced as a labelled estimate: the live lev-350m row (LiquidAI/LFM2.5-350M, in UNLISTED_BASE_MODELS, estimate 0.00462807881773399 per 1,000).
Measured p50 latency: 0.015 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: 0.18 s (est. incl. network (2x + 150 ms))
Self-hosted model: measured p50 0.015 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 = 0.18 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 5090 32 GB pod (France), offline after setup (egress cut and proven), author's custom code at the pinned commit via trust_remote_code from the sha256-verified local copy, bfloat16, no torch.compile