Open Jev alternatives on your own GPU

What is a decision model?

It doesn't write an answer. It chooses.

A decision model reads a text and answers typed questions: a choice among fixed options, yes or no with a probability, or a level on a scale. The answer comes back in a fraction of a second and always in the right format, so you can set thresholds on it. That suits ticket triage, risk flags and voice commands.

Which department gets this email?

choice · one choice out of three

Is the customer threatening to cancel?

noul · yes or no, with a probability

How urgent is the ticket?

score · 0, 1 or 2

Setup

Seven open projects, one reference and the same questions

Every model got exactly the same texts and questions in Jev's format. Jev was called through TypeSafe's API. The open models ran on their own engines on an RTX 3060 and an RTX 5060 Ti in the same machine, without Ollama or vLLM.

Model	Parameters	On disk	Format	Built on
Von	395M	3.0 GB	PyTorch	own
Laya	421M	1.5 GB	PyTorch	own
NanoJev	0.6B	2.3 GB	PyTorch	own
Kev 0.8B	0.8B	1.8 GB	bf16 + LoRA	Qwen3.5-0.8B-Base
Rizzo Flow 1.7B	1.7B	1.8 GB	GGUF q8_0	Spark-X2.5-1.7B
Rizzo Flow 4B	4B	4.4 GB	GGUF q8_0	Spark-X2.5-4B
Kev-4B	4B	9.0 GB	bf16 + LoRA	Qwen3.5-4B-Base
SemIf	4B	8.8 GB	bf16	Qwen3.5-4B
Nimble	9B	18 GB	bf16	Qwen3.5-9B
Jev	not published	–	cloud service	TypeSafe

Eight scenarios, 542 answers per model

Scenario	Language	What the model decides	Texts
support	English	Support email to one of three departments	20
types	English	Churn risk (yes/no) and urgency 0–2	16
intent	Swedish	Voice commands with speech recognition errors, nine intents	32
tree1	Swedish	Domain and action in one call	20
tree2	Swedish	Domain first, action in a second call	20
yesno	English	Four yes/no questions: cancel, angry, refund, data loss	32
yesno_crit	English	Same questions with criteria for yes and no	32
yesno_sv	Swedish	Same questions on Swedish texts	32

Results

Errors out of 542 answers. Fewer is better.

Without adaptation, Jev was best. Kev-4B was the best open model, and with two adapted questions it came within nine answers of Jev.

Jev95.2% correct26 errors

Kev-4B, adapted questions93.5% correct35 errors

Voting: Kev-4B, Rizzo 4B, Laya91.7% correct45 errors

Nimble 9B, adapted questions91.1% correct48 errors

Kev-4B91.0% correct49 errors

Nimble 9B88.6% correct62 errors

Rizzo 4B88.0% correct65 errors

Laya, adapted questions87.1% correct70 errors

Rizzo 1.7B83.0% correct92 errors

Kev 0.8B77.1% correct124 errors

Laya73.4% correct144 errors

TypeSafe JevOpen model on our own GPU

Voting: majority for choices, mean for yes/no and scores. Laya loses 68 answers when the yes/no questions have criteria.

Per scenario, without adaptation

Scenario	Jev	Kev-4B	Nimble 9B	Rizzo 4B	Rizzo 1.7B	Kev 0.8B	Laya
support /20	20	20	19	19	18	20	13
types /16	13	11	12	10	13	9	10
intent /32	32	26	27	22	19	20	24
tree1 /20	18	17	17	15	12	11	14
tree2 /20	18	17	17	15	11	11	14
yesno /128	121	118	113	115	114	95	119
yesno_crit /128	126	121	118	123	114	112	60
yesno_sv /128	120	116	109	114	111	97	98

Fully correct cases per scenario. For the yes/no scenarios: correct answers out of 128 (32 texts × 4 questions).

How we got closer

Two changes. No training.

Same model, same weights. We only changed how the questions were asked. Each change was chosen on half of the texts and checked on the other half.

Label the text

Every voice command got the prefix “Swedish voice command (may contain speech recognition errors):”.

Move one threshold

The question “is the customer angry?” counted as yes only at probability 0.94 instead of 0.5.

Scenario	Jev	Kev-4B	Kev-4B adapted	Nimble 9B adapted	Laya adapted
support	20/20	20/20	20/20	19/20	13/20
types	29/32	27/32	27/32	28/32	26/32
intent	32/32	26/32	29/32	31/32	28/32
tree1	35/37	34/37	36/37	35/37	29/37
tree2	33/37	31/37	35/37	33/37	28/37
yesno	121/128	118/128	118/128	113/128	121/128
yesno_crit	126/128	121/128	126/128	126/128	121/128
yesno_sv	120/128	116/128	116/128	109/128	106/128
Total of 542	516	493	507	494	472

Correct answers per scenario. Nimble 9B got the same two changes, with threshold 0.98. Laya got different changes: yes/no questions without criteria, Swedish questions and labelling for Swedish yes/no texts, and threshold 0.81 for the angry question.

Swedish and size

Swedish shows the biggest gap. Bigger wasn't better.

32/32

Jev on Swedish voice commands. Best open models with adapted questions: Nimble 9B 31, Kev-4B 29.

49 / 62

Errors for Kev-4B (4B, one card) against Nimble 9B (9B, two cards), both without adaptation.

421M

Laya, the smallest in the test, was second-best locally on Swedish voice commands and answers in 16 ms.

Swedish voice commands against response time

Correct out of 32 on the y-axis, median ms on the x-axis. Up and to the left is best.

Without adapted questions. Local models on an RTX 3060, Nimble 9B across both cards. Jev's time includes the network.

Speed

Every local model answers faster than Jev

Median per support email. The goal was the most correct answers, not the lowest time, but time decides where a model is usable.

Laya16 ms

Kev 0.8B19 ms

Von19 ms

NanoJev33 ms

Rizzo 1.7B40 ms

Rizzo 4B81 ms

Kev-4B87 ms

SemIf 4B122 ms

Nimble 9B †186 ms

Jev *230 ms

Local models on an RTX 3060. *Jev including the network. †Across an RTX 5060 Ti and 3060, without fast kernels.

RTX 3060 against RTX 5060 Ti

RTX 3060 12 GBRTX 5060 Ti 16 GB

Laya16 ms12 ms · 1.3×

Kev 0.8B19 ms12 ms · 1.6×

Rizzo 1.7B40 ms25 ms · 1.6×

Rizzo 4B81 ms50 ms · 1.6×

Kev-4B87 ms74 ms · 1.2×

SemIf 4B122 ms65 ms · 1.9×

Same answers on both cards except two borderline cases. Kev-4B ran on the 5060 Ti without its fastest kernel, which couldn't be built for the card with our CUDA version.

RTX 3060 with unlocked BF16/FP16

With the unlock, Kev-4B got 18–27% faster on the 3060 with identical answers, on par with the 5060 Ti.

3060 before3060 unlocked5060 Ti

Support email87 ms71 ms · −18%74 ms

Decision tree126 ms92 ms · −27%84 ms

Yes/no with criteria129 ms94 ms · −27%86 ms

Conclusions

Kev-4B gets close, with the right questions

Best local: Kev-4B

507 of 542 correct with adapted questions, on one 12 GB card. Beats Jev on the decision trees and matches it on yes/no with criteria.

Questions move the most

Phrasing and thresholds gained more correct answers than both larger models and GPU settings.

Still open: Swedish

Voice commands and Swedish customer emails are where Jev still leads. Next steps are labelling for customer emails and fine-tuning on Swedish examples.

If you need fast, structured decisions and want to keep your data in-house, the open models go a long way. If you need breadth and Swedish without extra work, Jev is still the best.

Method and caveats

Small test sets, stable answers

16–32 texts per scenario

One case is worth 3–6 percentage points. Differences of one or two cases are noise.

Deterministic answers

Five runs per model gave identical answers. Different machines and graphics cards differed by at most two answers.

Selection against a control half

The adaptations were chosen on half of the texts and checked on the rest. Confirm on more texts before you build on them.

Jev's times include the network, since Jev was called as a cloud service. The local times are pure GPU time. The models come from DataCamp's list of open Jev alternatives.

Glossary

Jev: TypeSafe's decision model, called as a cloud service. The reference every model is compared against.
Parameters (M, B): The model's learned numbers. M = million, B = billion. More parameters need more memory.
Base model: The general language model a decision model is built on, here often Qwen3.5.
LoRA: A small add-on file (adapter) that adapts a base model to a task without changing the whole model.
bf16, fp16: 16-bit number formats for the weights. Standard for GPU inference, not quantization.
Quantization: Weights stored with fewer bits, such as 8 or 4, to save memory.
GGUF, q8_0: The llama.cpp file format, and 8-bit quantization. Rizzo Flow ships that way.
llama.cpp: An inference engine for language models written in C++, the one Ollama builds on.
choice, noul, score: The question types: fixed options, yes/no with a probability, and ordered levels such as urgency 0–2.
criteria: Descriptions of what each option, or yes and no, means. Helps Kev, hurts Laya.
Threshold: The probability at which an answer counts as yes. Default 0.5.
Control half: Half of the texts, used only to check a setting, never to choose it.
Fused kernels: Several small GPU programs merged into one. Makes Kev-4B 12–19% faster on an RTX 3060.
VRAM: The graphics card's own memory: 12 GB on the RTX 3060, 16 GB on the RTX 5060 Ti.