d1-omni:600m
d1-omni · Liquid AI · 587M · ONNX Runtime · LFM Open License v1.0
- Score
- 45.5
- 95 % interval 43.5–47.7
- Rank
- 34
- of 66 scored models
- Public benchmark
- 63.9
- macro accuracy, %
- typed-decisions
- 57.5
- accuracy, %
By task
Chance-corrected accuracy on each of the 14 tasks: 0 is guessing, 100 is every answer right.
Workflows
typed-decisions
37.7
Classification
MASSIVE en
82.8
MASSIVE de
70.0
Civil Comments
72.0
Aegis 2
28.0
Understanding
MultiNLI
69.9
VitaminC
48.9
PAWS
40.8
Knowledge
BoolQ
38.0
SQuAD 2
53.2
PubMedQA
46.6
Judging
HelpSteer 2
27.7
SummEval rel.
3.1
SummEval cons.
18.4
Accuracy and calibration
| Benchmark | Accuracy | Choice | Noul | Score | ECE | Brier | Rejected |
|---|---|---|---|---|---|---|---|
| Public benchmark | 63.9(62.5–65.3) | 73.1 | 73.2 | 33.1 | 0.075 | 0.446 | 0 / 3,880 |
| typed-decisions | 57.5(55.3–59.7) | 60.7 | 70.8 | 45.0 | 0.066 | 0.537 | 0 / 400 requests |
Accuracy in %, with its 95 % bootstrap interval. A rejected request (the model can’t take the question, or the input is too long for it) counts as wrong.
Speed and memory
| Device | p50 | p90 | Load | VRAM | RAM peak | Precision |
|---|---|---|---|---|---|---|
| RTX 5090 CUDA | 16.6 ms | 17.4 ms | 7.90 s | 2.6 GB | 1.9 GB | F32 |
| Threadripper 3960X CPU | 960 ms | 992 ms | 5.80 s | – | 1.8 GB | F32 |
Five-question requests over HTTP, one at a time, each a different short message. VRAM: the rise of the GPU’s used memory while the model is loaded. RAM peak: the runner process’s peak resident memory.
Parity with the authors’ code
| Device | Same decision | Max logit diff. | Result | Record |
|---|---|---|---|---|
| RTX 4090 CPU | 456/456 | 1.4e-4 | pass | results/runs/2026-10-24-parity-rtx4090-cpu.json |
| RTX 4090 CUDA | 456/456 | 8.6e-5 | pass | results/runs/2026-10-24-parity-rtx4090-cuda.json |
Reproduce
The tools are in convert/ollaya_convert and need only Python’s standard library. Run them against a local Ollaya 0.13.0 with the same model digest.
ollaya pull d1-omni:600m
# typed-decisions; export the test split once with
# uv run python -m ollaya_convert.bench_typed export --out typed.jsonl
python -m ollaya_convert.bench_typed run --model d1-omni:600m \
--url http://127.0.0.1:11435 --data typed.jsonl --out typed/d1-omni-600m
# Bespoke Labs' public benchmark; build its 13 subsets first
NIMBLE_SRC=nimble python -m ollaya_convert.bench_public run --model d1-omni:600m \
--url http://127.0.0.1:11435 --data public --out public/d1-omni-600m
# speed and memory
python -m ollaya_convert.bench_latency --models d1-omni:600m --memory \
--machine my-machine --device cuda --out latency.json