verdict-models/OneJev1.6

OneJev 1.6 (Verdict-ready)

The OneJev 1.6 family of multimodal System One decision models — text, image, and video inputs — packaged for Verdict, a high-performance serving engine for System One decision models.

This repository is a variant family: four variants — 0.8b, 4b, 9b, 27b — each a complete, self-contained package. Pick one with the :variant syntax; the bare name lists the variants and exits instead of guessing. One command to serve — the 9b variant, for example:

# GPU host
docker run --rm --gpus all -p 8080:8080 \
    -v ~/.cache/verdict:/root/.cache/verdict \
    sgyaqing/verdict:latest-cuda serve verdict-models/OneJev1.6:9b --host 0.0.0.0

# ModelScope:
docker run --rm --gpus all -p 8080:8080 \
    -e VERDICT_USE_MODELSCOPE=1 \
    -v ~/.cache/verdict:/root/.cache/verdict \
    sgyaqing/verdict:latest-cuda serve sgyaqing/OneJev1.6:9b --host 0.0.0.0

Endpoints: the standard decision protocols POST /v1/systemone and POST /v1/decisions, native POST /v1/choice | /v1/score | /v1/noul, plus an OpenAI-compatible layer (/v1/chat/completions, /v1/models).

Files

File	Purpose
variants.json	the variant index
<variant>/model.json	the manifest — budgets, and the declaration of which official repo holds the checkpoint
SHA256SUMS	checksums for the files above

No weights ship here. On first serve Verdict downloads the checkpoint (weights, configs, tokenizer) from the official OmniJev repo named in the variant's manifest, into the local cache.

Provenance & fidelity

Each variant serves its official checkpoint in bf16 (its own precision), verified against the golden standard (the official OneJev SDK on CPU fp32, 31-case corpus per tier): zero answer flips on CUDA (RTX 4090).

Accuracy PK — against the model author's own server (qev 0.1.4); both servers on one RTX 4090, graded against the same golden standard. Worst-case deviation across the corpus:

Variant	Verdict	qev	result
0.8b	2.49e-2	8.55e-1	Verdict ahead
4b	4.95e-2	9.03e-1	Verdict ahead
9b	4.59e-2	8.99e-1	Verdict ahead

Performance

Measured end to end over HTTP against the model author's own server (qev 0.1.4) — each cell a 10-second duration window after a warm matrix: 0.8b/4b/9b on one RTX 4090 (48 GB), 27b on one NVIDIA A100 (80 GB). A ratio above 1x means Verdict is ahead.

0.8b

workload	concurrency	Verdict rps	qev rps	rps ratio	Verdict P50 / P99 (ms)	qev P50 / P99 (ms)	P50 / P99 ratio
text	1	116.7	87.5	1.33x	8.90 / 9.30	11.37 / 11.93	1.28x / 1.28x
	8	245.7	89.9	2.73x	33.16 / 35.56	89.42 / 91.05	2.70x / 2.56x
	64	289.5	89.6	3.23x	225.43 / 228.53	693.37 / 928.39	3.08x / 4.06x
image	1	37.5	38.0	0.99x	26.86 / 28.70	26.31 / 28.73	0.98x / 1.00x
	8	135.6	38.4	3.53x	59.09 / 60.98	205.37 / 422.47	3.48x / 6.93x
	64	141.2	44.8	3.15x	472.29 / 479.76	1639.10 / 1697.90	3.47x / 3.54x
video	1	30.1	26.8	1.12x	32.24 / 40.02	31.30 / 35.30	0.97x / 0.88x
	8	109.2	31.5	3.47x	72.57 / 83.08	251.40 / 464.83	3.46x / 5.59x
	64	115.2	32.0	3.60x	583.80 / 608.71	2018.50 / 2143.50	3.46x / 3.52x

4b

workload	concurrency	Verdict rps	qev rps	rps ratio	Verdict P50 / P99 (ms)	qev P50 / P99 (ms)	P50 / P99 ratio
text	1	38.1	31.3	1.22x	25.98 / 27.01	31.99 / 32.38	1.23x / 1.20x
	8	47.3	32.3	1.46x	163.49 / 185.22	253.52 / 254.76	1.55x / 1.38x
	64	55.2	32.0	1.72x	1275.20 / 1425.70	2017.70 / 2243.00	1.58x / 1.57x
image	1	17.2	16.8	1.02x	57.51 / 60.47	57.68 / 60.92	1.00x / 1.01x
	8	51.6	17.6	2.93x	156.16 / 160.33	461.55 / 706.21	2.96x / 4.41x
	64	56.8	19.2	2.96x	1254.90 / 1267.90	4074.00 / 4076.10	3.25x / 3.21x
video	1	15.2	13.2	1.15x	65.60 / 73.36	64.40 / 71.10	0.98x / 0.97x
	8	45.3	16.0	2.83x	176.36 / 239.65	512.21 / 572.87	2.90x / 2.39x
	64	51.2	19.2	2.67x	1409.70 / 1514.20	4451.10 / 4452.30	3.16x / 2.94x

9b

workload	concurrency	Verdict rps	qev rps	rps ratio	Verdict P50 / P99 (ms)	qev P50 / P99 (ms)	P50 / P99 ratio
text	1	23.4	20.7	1.13x	42.51 / 43.85	48.48 / 48.85	1.14x / 1.11x
	8	28.8	20.8	1.38x	282.91 / 287.41	385.88 / 397.18	1.36x / 1.38x
	64	34.4	25.6	1.34x	2245.00 / 2381.70	3091.00 / 3135.50	1.38x / 1.32x
image	1	11.7	9.6	1.22x	81.42 / 83.24	79.90 / 2471.30	0.98x / 29.69x
	8	35.4	12.8	2.77x	227.51 / 304.12	645.06 / 690.39	2.84x / 2.27x
	64	41.2	12.8	3.22x	1819.90 / 1823.80	5496.60 / 5497.30	3.02x / 3.01x
video	1	11.2	10.7	1.05x	88.36 / 93.97	87.91 / 92.22	0.99x / 0.98x
	8	31.4	11.2	2.80x	247.69 / 339.28	711.63 / 936.82	2.87x / 2.76x
	64	38.0	12.8	2.97x	1974.30 / 2126.60	5793.00 / 5793.50	2.93x / 2.72x

27b (A100 80 GB)

workload	concurrency	Verdict rps	qev rps	rps ratio	Verdict P50 / P99 (ms)	qev P50 / P99 (ms)	P50 / P99 ratio
text	1	9.8	8.5	1.15x	102.31 / 103.45	117.76 / 120.00	1.15x / 1.16x
	8	12.9	9.1	1.42x	640.28 / 736.38	940.88 / 942.44	1.47x / 1.28x
	64	17.6	12.8	1.37x	5095.90 / 5592.40	8209.50 / 8210.10	1.61x / 1.47x
image	1	5.7	2.5	2.28x	178.32 / 180.22	177.50 / 5888.90	1.00x / 32.7x
	8	14.3	5.6	2.55x	555.69 / 729.59	1416.60 / 1529.50	2.55x / 2.10x
	64	20.0	n/a †	—	4457.90 / 4884.30	—	—
video	1	5.3	5.4	0.98x	188.07 / 208.21	184.72 / 190.44	0.98x / 0.91x
	8	14.4	6.0	2.40x	574.03 / 579.80	1487.60 / 1496.60	2.59x / 2.58x
	64	19.5	n/a †	—	4572.40 / 4861.90	—	—

At image and video `c=64`, `qev` completed **zero** responses inside the 10-second window: under sustained load its queue bleeds over from the preceding cell and never drains, so every request was still in flight at the deadline (the per-request timeout is 60 s — nothing errored). Those `qev` cells are therefore reported as n/a, not as a measured zero. For reference, a supplementary run from a drained queue with an extended per-request timeout collected 23 and 8 responses (8.7 and 7.2 rps; P50/P99 2883.3/4128.1 and 1523.2/1523.3 ms) — a recovery-condition estimate, not the sustained-load protocol result, so no ratios are derived from it. `qev`'s image `c=1` P99 reflects a stall inside its own window (its P50 there is 177.5 ms).

License

The models are Apache-2.0, as declared by OmniJev; credit for the models belongs to their authors. This repository holds only the Verdict manifests.