Best Decision Models That Run on a CPU

CPU-class decision models (encoders and small decoders, under about 1.5B parameters) run next to the data with no GPU. Models are ranked by their Full score on the Jev Decision Index, a third-party board of 37 benchmarks with private held-out sets, until scores from our own Metatext Decision Benchmark are published. Latency is the board's median per call on its GPU runner.

Ranked list (top 25)

#	Model	Decision Index	ECE	p50 latency	Per 1k decisions	Params
1	vLLM-SR d3-lite (0.8B) vLLM Semantic RouterOpen weights	32.3	0.094	37 ms	—	853M
2	OneJev1.5-0.8B OmniJevOpen weights	29.0	0.035	83 ms	—	1B
3	jiwo 0.8B v2 eljiwoOpen weights	24.9	0.068	36 ms	—	752M
4	EXAONE-4.0-1.2B-JEV v0.3 carrtesyOpen weights	23.3	0.079	149 ms	—	1B
5	Sieve 0.8B Sthānika AIOpen weights	23.1	0.088	52 ms	—	590M
6	Crivo-150M tardellirsOpen weights	22.2	0.143	13 ms	—	150M
7	Saina Helm 2.0 SainaOpen weights	20.8	0.247	39 ms	—	853M
8	vLLM-SR d3-edge (0.6B) vLLM Semantic RouterOpen weights	20.2	0.141	19 ms	—	594M
9	JPT-0.8B kirpOpen weights	18.7	0.069	118 ms	—	873M
10	Sifr 0.8B v3.1 mohamedlotfy50Open weights	17.8	0.120	35 ms	—	870M
11	tasksource Decider nano v4 (150M) tasksourceOpen weights	17.6	0.069	14 ms	—	149M
12	vLLM-SR Decision 2.0 Eos 0.8B vLLM Semantic RouterOpen weights	17.5	0.130	25 ms	—	870M
13	DecisionTune 1.0 DecisionTuneOpen weights	16.7	0.147	14 ms	—	395M
14	Bekko System One v0 400M hotchpotchOpen weights	15.2	0.111	17 ms	—	400M
15	Dinah-0 LukitaduarteOpen weights	13.9	0.104	14 ms	—	150M
16	Kev 0.8B r15 Jared PalmerOpen weights	13.7	0.040	45 ms	—	870M
17	vLLM-SR Decision 2.0 Kai 0.6B vLLM Semantic RouterOpen weights	13.6	0.038	19 ms	—	600M
18	Tev1-0.8B-experimental TogetherOpen weights	12.0	0.126	29 ms	—	873M
19	Jev-Style 0.8B v3 chaoliangUNSWOpen weights	11.9	0.084	36 ms	—	870M
20	Crivo-32M tardellirsOpen weights	11.9	0.087	7 ms	—	32M
21	Intern-Decision-0.8B Intern Large ModelsOpen weights	11.8	0.025	34 ms	—	873M
22	vera-spark scar-aiOpen weights	11.5	0.157	80 ms	—	638M
23	GLiNER2.5-Decide FastinoOpen weights	9.2	0.088	23 ms	—	486M
24	LiquidAI d1-omni-600M Liquid AIOpen weights	9.1	0.128	11 ms	—	587M
25	Bosun v3.1 0.6B Hanno LabsOpen weights	8.7	0.142	39 ms	—	596M

Scores, calibration (ECE) and latency from the Jev Decision Index; cost per 1,000 decisions at 200 input tokens each. Updated daily.

More decision-model rankings

Best decision model Best calibrated Cheapest Open weights Jev alternatives

Frequently asked questions

What is the best decision model that runs on a CPU?

vLLM-SR d3-lite (0.8B) ranks first among CPU-class decision models on the Jev Decision Index.

How accurate are CPU decision models compared with large ones?

They trail the 27B-class leaders by a wide margin on the Decision Index, but answer in milliseconds on commodity hardware, which suits high-volume routing and filtering.