Decision models, ranked
A decision model does not write text. It picks one option, gives a score, or answers yes or no with a probability, in a fraction of a second. TypeSafe AI released the first one, Jev, in September 2026; dozens of open models now copy the idea.
Models ranked173147 with a Hugging Face page
Best open model69.1StartLux-Decision v1.1 27B
Jev 1.14 (TypeSafe)61.5closed model, the reference
Beat Jev12open models score higher
Full ranking
| Rank | Model | Decision Index | vs Jev | Type | Size | Download | 4-bit needs | Time per decision | Calibration error |
|---|---|---|---|---|---|---|---|---|---|
| 1 | StartLux-Decision v1.1 27B StartLux · based on Qwen3.8-27B |
69.1 | +7.6 | Fine-tuned model | 27.8 B | Yes | 18 GB | 80 ms | 2.6% |
| 2 | vLLM-SR d3 (27B) vLLM Semantic Router · based on Qwen3.8-27B |
66.3 | +4.7 | Fine-tuned model | 26.1 B | Yes | 17 GB | 97 ms | 4.3% |
| 3 | linne-oak 1.0 (31B) MISA-AICORE · based on gemma-4-31B-it |
65.6 | +4.1 | Fine-tuned model | 31.3 B | Yes | 20 GB | 105 ms | 10.5% |
| 4 | Perplexity Decider v1.2 (27B) Perplexity · based on Qwen3.8-27B |
65.5 | +4.0 | Fine-tuned model | 26.1 B | Yes | 17 GB | 144 ms | 5.5% |
| 5 | Decider 31B (NVFP4) · based on gemma-4-31B-it |
64.6 | +3.1 | Fine-tuned model | 32.7 B | Yes | 21 GB | 72 ms | 2.4% |
| 6 | Amos 1 Large (27B) · based on Qwen3.8-27B |
64.1 | +2.5 | Fine-tuned model | 27.8 B | Yes | 18 GB | 148 ms | 3.2% |
| 7 | Sieve 31B Sthānika AI · based on gemma-4-31B-it |
63.8 | +2.3 | LoRA adapter | 31.3 B | Yes | 20 GB | 253 ms | 3.3% |
| 8 | Lychee-27B · based on Qwen3.8-27B |
63.0 | +1.5 | Fine-tuned model | 26.1 B | Yes | 17 GB | 97 ms | 10.8% |
| 9 | KnowLine 35B-A3B Gen1 PelaAI · based on Qwen3.5-35B-A3B |
62.9 | +1.4 | LoRA adapter | 36 B | Yes | 23 GB | 41 ms | 4.8% |
| 10 | Bobcat 1.3 (31B) · based on gemma-4-31B-it |
62.8 | +1.3 | LoRA adapter | 31.3 B | Yes | 20 GB | 64 ms | 10.2% |
| 11 | span-02 (Gemma 4 26B-A4B) · based on gemma-4-26B-A4B-it |
62.6 | +1.0 | Fine-tuned model | 25.8 B | Yes | 17 GB | 33 ms | 11.4% |
| 12 | Darwin-27B-ZTC FINAL_Bench · based on Darwin-27B-RSI |
62.5 | +0.9 | Fine-tuned model | 25.6 B | Yes | 17 GB | 104 ms | 4.5% |
| 13 | MORJ 1.0.1 (Gemma 4 31B) · based on gemma-4-31B-it |
61.0 | -0.6 | Technique on a stock model | 31.3 B | No | 20 GB | 202 ms | 9.2% |
| 14 | AnyJev L0 · Gemma-4-31B · based on gemma-4-31B-it |
60.2 | -1.3 | Technique on a stock model | 31.3 B | No | 20 GB | 98 ms | 24.0% |
| 15 | Fastino GLiDE no-thinking (28B) Fastino |
60.2 | -1.3 | Fine-tuned model | 27.8 B | Open weights coming soon | 18 GB | 98 ms | 4.0% |
| 16 | Sieve 27B Sthānika AI · based on Qwen3.8-27B |
60.2 | -1.4 | LoRA + decision head | 27.8 B | Yes | 18 GB | 101 ms | 6.5% |
| 17 | Jev 1.13 | 60.1 | -1.4 | Hosted API | — | No | — | 524 ms | 7.8% |
| 18 | Torchcast Decision 27B Torchcast AI · based on Qwen3.8-27B |
59.9 | -1.6 | LoRA adapter | 27.8 B | Yes | 18 GB | 100 ms | 2.4% |
| 19 | OneJev1.5-27B OmniJev · based on Qwen3.8-27B |
59.9 | -1.6 | Fine-tuned model | 27.4 B | Yes | 18 GB | 224 ms | 2.4% |
| 20 | Shisa DE-2 (repeat) Shisa.AI · based on gemma-4-26B-A4B-it |
59.8 | -1.8 | LoRA adapter | 25.8 B | Yes | 17 GB | 57 ms | 20.7% |
| 21 | Matilda Jev v1.5 Maincode · based on Qwen3.8-27B |
59.2 | -2.3 | Fine-tuned model | 26.1 B | Yes | 17 GB | 89 ms | 7.7% |
| 22 | deck31b (Gemma 4 31B) · based on gemma-4-31B-it |
59.0 | -2.5 | Technique on a stock model | 32.7 B | No | 21 GB | 219 ms | 20.4% |
| 23 | Kev 27B · based on Qwen3.8-27B |
58.8 | -2.8 | Fine-tuned model | 27.8 B | Yes | 18 GB | 135 ms | 2.2% |
| 24 | Quyet-1.0-Large · based on gemma-4-31B-it |
58.7 | -2.8 | LoRA adapter | 32.7 B | Yes | 21 GB | 105 ms | 7.8% |
| 25 | RSI-Jev v6.1 VL 27B (effort auto) · based on Qwen3.8-27B |
58.7 | -2.9 | LoRA adapter | 27.8 B | Yes | 18 GB | 76 ms | 7.7% |
| 26 | decisio 0.6.0 (Gemma 4 31B) · based on gemma-4-31B-it |
58.5 | -3.0 | Technique on a stock model | 31.3 B | No | 20 GB | 54 ms | 14.2% |
| 27 | Qwen Decision 27B noncausal LoRA · based on Qwen3.8-27B |
58.5 | -3.0 | LoRA + decision head | 27.8 B | Yes | 18 GB | 127 ms | 5.1% |
| 28 | garas (Gemma 4 31B) · based on gemma-4-31B-it |
58.3 | -3.2 | LoRA adapter | 31.3 B | Yes | 20 GB | 117 ms | 4.9% |
| 29 | Wald-26B-A4B ORG2 AI · based on gemma-4-26B-A4B-it |
58.3 | -3.2 | Fine-tuned model | 25.8 B | Yes | 17 GB | 90 ms | 4.3% |
| 30 | linne-fern 1.0 (27B) MISA-AICORE · based on Qwen3.8-27B |
57.8 | -3.7 | Fine-tuned model | 27.8 B | Yes | 18 GB | 108 ms | 7.9% |
| 31 | Blink v0.3 26B-A4B (NVFP4) Pixilab.ai · based on gemma-4-26B-A4B-it |
57.8 | -3.8 | Fine-tuned model | 25.8 B | Yes | 17 GB | 25 ms | 4.0% |
| 32 | Decider chat · Gemma-4-31B · based on gemma-4-31B |
57.6 | -3.9 | Technique on a stock model | 32.7 B | No | 21 GB | 109 ms | 4.7% |
| 33 | SynACK Decide 26B-A4B v7 · based on gemma-4-26B-A4B-it |
57.5 | -4.1 | LoRA adapter | 25.8 B | Yes | 17 GB | 129 ms | 6.9% |
| 34 | Surogate Rune 26B-A4B v3 Surogate · based on gemma-4-26B-A4B-it |
57.4 | -4.1 | Fine-tuned model | 25.8 B | Yes | 17 GB | 121 ms | 12.0% |
| 35 | Xor 26B-A4B v1.1 Juspay Technologies · based on gemma-4-26B-A4B-it |
57.1 | -4.5 | LoRA adapter | 25.8 B | Yes | 17 GB | 110 ms | 5.8% |
| 36 | JEV-35B AutoTrust AI Lab · based on Qwen3.5-35B-A3B |
57.0 | -4.6 | LoRA + decision head | 36 B | Yes | 23 GB | 51 ms | 8.5% |
| 37 | AJev lora1 (Gemma 4 26B-A4B) · based on gemma-4-26B-A4B-it |
56.4 | -5.1 | LoRA adapter | 25.8 B | Yes | 17 GB | 89 ms | 4.8% |
| 38 | GEV-26B-Decide AutoTrust AI Lab · based on gemma-4-26B-A4B-it |
56.2 | -5.4 | LoRA adapter | 25.8 B | Yes | 17 GB | 129 ms | 7.2% |
| 39 | Nace.AI Drex v2.0 (35B-A3B) Nace.AI · based on Qwen3.6-35B-A3B |
55.9 | -5.6 | Fine-tuned model | 35.1 B | Yes | 22 GB | 276 ms | 1.8% |
| 40 | frontier-infra Jebadiah 27B v2.1 Frontier Infra · based on Qwen3.8-27B |
55.9 | -5.6 | LoRA adapter | 27.8 B | Yes | 18 GB | 110 ms | 6.1% |
| 41 | vLLM-SR Decision 2.0 Vega 27B vLLM Semantic Router · based on Qwen3.8-27B |
55.9 | -5.6 | LoRA + decision head | 27.8 B | Yes | 18 GB | 120 ms | 8.6% |
| 42 | vLLM-SR d3-flash (9B) vLLM Semantic Router · based on Qwen3.5-9B |
55.8 | -5.7 | Fine-tuned model | 8.4 B | Yes | 6.3 GB | 54 ms | 5.6% |
| 43 | JEV-27B AutoTrust AI Lab · based on Qwen3.8-27B |
55.6 | -5.9 | Fine-tuned model | 27.8 B | Yes | 18 GB | 102 ms | 7.6% |
| 44 | JADE · based on Qwen3.8-27B |
55.6 | -6.0 | LoRA adapter | 27.8 B | Yes | 18 GB | 133 ms | 23.8% |
| 45 | Eikos-27B · based on Qwen3.8-27B |
55.3 | -6.2 | LoRA adapter | 27.8 B | Yes | 18 GB | 130 ms | 5.7% |
| 46 | Perplexity Decider Flash v1.2 (4B) Perplexity · based on Qwen3.5-4B |
54.8 | -6.7 | Fine-tuned model | 4.5 B | Yes | 3.9 GB | 75 ms | 5.2% |
| 47 | reflex Qwen3.8-27B-FP8 (pmscr) · based on Qwen3.8-27B-FP8 |
54.7 | -6.8 | Technique on a stock model | 27.8 B | No | 18 GB | 149 ms | 4.5% |
| 48 | Bespoke Nimble 9B v3 Bespoke Labs · based on Qwen3.5-9B |
54.7 | -6.9 | LoRA adapter | 9.7 B | Yes | 7.1 GB | 74 ms | 12.3% |
| 49 | OneJev1.5-9B OmniJev · based on Qwen3.5-9B |
53.3 | -8.2 | Fine-tuned model | 9.4 B | Yes | 6.9 GB | 111 ms | 2.9% |
| 50 | Cloudflare clef Cloudflare · based on Qwen3.8-27B |
53.1 | -8.5 | Fine-tuned model | 27.8 B | Yes | 18 GB | 102 ms | 4.0% |
| 51 | vLLM-SR d3-mini (4B) vLLM Semantic Router · based on Qwen3.5-4B |
52.6 | -8.9 | Fine-tuned model | 4.5 B | Yes | 3.9 GB | 50 ms | 5.6% |
| 52 | simple-jev · Qwen3.8-27B (featherless) · based on Qwen3.8-27B |
51.8 | -9.7 | Technique on a stock model | 27.8 B | No | 18 GB | 373 ms | 11.3% |
| 53 | JPT-35B-A3B · based on Qwen3.5-35B-A3B |
51.4 | -10.1 | LoRA adapter | 36 B est. | Yes | 23 GB | 250 ms | 4.9% |
| 54 | Decider chat · Qwen3.6-27B · based on Qwen3.6-27B |
51.3 | -10.3 | Technique on a stock model | 27.8 B | No | 18 GB | 84 ms | 2.1% |
| 55 | Xor 1.2 Juspay Technologies · based on Qwen3.6-35B-A3B |
51.1 | -10.4 | LoRA adapter | 36 B | Yes | 23 GB | 150 ms | 2.4% |
| 56 | Sieve 9B v2 Sthānika AI · based on Qwen3.5-9B |
50.9 | -10.6 | LoRA + decision head | 9.7 B | Yes | 7.1 GB | 51 ms | 7.3% |
| 57 | AJev lora7 (Gemma 4 12B) · based on gemma-4-12B-it |
50.7 | -10.9 | LoRA adapter | 12 B | Yes | 8.5 GB | 69 ms | 8.2% |
| 58 | Nace.AI Drex v1.5 Nace.AI · based on MiMo-V2.6-Distill-Qwen-9B |
49.7 | -11.8 | Fine-tuned model | 8.9 B | Yes | 6.6 GB | 54 ms | 9.0% |
| 59 | Standard One 8B SH (v4) Standard Thinking · based on StandardOne-8B |
49.6 | -12.0 | LoRA + decision head | 8.9 B | Yes | 6.6 GB | 90 ms | 10.7% |
| 60 | Winnow-12B · based on gemma-4-12B |
49.3 | -12.3 | LoRA adapter | 12 B | Yes | 8.5 GB | 73 ms | 16.8% |
| 61 | OneJev1.5-4B OmniJev · based on Qwen3.5-4B |
49.2 | -12.4 | Fine-tuned model | 5.2 B | Yes | 4.4 GB | 113 ms | 3.4% |
| 62 | CoCo-Decision-4B 2.3.0 Corners · based on Qwen3.5-4B |
48.8 | -12.7 | LoRA + decision head | 4.2 B | Yes | 3.7 GB | 63 ms | 11.6% |
| 63 | Decider 35B-A3B · based on Qwen3.5-35B-A3B-Base |
48.5 | -13.0 | Fine-tuned model | 36 B | Yes | 23 GB | 101 ms | 2.3% |
| 64 | JevK5 9B v0.3.3 · based on Qwen3.5-9B |
48.4 | -13.2 | LoRA adapter | 9.7 B | Yes | 7.1 GB | 40 ms | 6.0% |
| 65 | Cloudflare clef-omni (30B-A3B) Cloudflare · based on Qwen3-Omni-30B-A3B-Instruct |
48.3 | -13.3 | Fine-tuned model | 31.8 B | Yes | 20 GB | 191 ms | 4.0% |
| 66 | vLLM-SR Decision 2.0 Lux 9B vLLM Semantic Router · based on Qwen3.5-9B |
48.2 | -13.3 | Fine-tuned model | 9.7 B | Yes | 7.1 GB | 37 ms | 5.6% |
| 67 | decisio 0.4.0 (Qwen3.6-35B-A3B) · based on Qwen3.6-35B-A3B |
47.7 | -13.8 | Technique on a stock model | 36 B | No | 23 GB | 71 ms | 5.5% |
| 68 | Cloudflare clef-flash Cloudflare · based on Qwen3.5-9B |
47.6 | -13.9 | Fine-tuned model | 9.7 B | Yes | 7.1 GB | 53 ms | 2.5% |
| 69 | decisio 0.4.0 (Gemma 4 12B) · based on gemma-4-12B-it |
47.5 | -14.1 | Technique on a stock model | 12 B | No | 8.5 GB | 38 ms | 10.3% |
| 70 | JoshuaSP diffusiongemma (open-jev) · based on diffusiongemma-26B-A4B-it |
47.3 | -14.2 | Technique on a stock model | 25.8 B | No | 17 GB | 260 ms | 21.6% |
| 71 | Hopper 12B (trained) HopitAI · based on gemma-4-12B-it |
47.3 | -14.3 | LoRA adapter | 12 B | Yes | 8.5 GB | 290 ms | 7.6% |
| 72 | ezjev 4B s2 · based on Qwen3.5-4B |
47.0 | -14.6 | LoRA adapter | 4.7 B | Yes | 4.0 GB | 36 ms | 6.3% |
| 73 | PurpleMIST-Flash-1.0 Olaverse Lab · based on Qwen3.5-9B-Base |
46.8 | -14.7 | LoRA + decision head | 7.9 B | Yes | 6.0 GB | 53 ms | 3.2% |
| 74 | RSI-Jev v6.1 VL 4B · based on Qwen3.5-4B-Base |
46.3 | -15.2 | Fine-tuned model | 4.7 B | Yes | 4.0 GB | 58 ms | 4.9% |
| 75 | Sieve 4B Sthānika AI · based on Qwen3.5-4B |
46.1 | -15.4 | LoRA + decision head | 4.7 B | Yes | 4.0 GB | 93 ms | 3.1% |
| 76 | Jevfire · based on Qwen3.8-27B |
45.5 | -16.0 | Technique on a stock model | 27.8 B | No | 18 GB | 90 ms | 5.2% |
| 77 | JEV-9B AutoTrust AI Lab · based on Qwen3.5-9B |
45.2 | -16.3 | Fine-tuned model | 9.7 B | Yes | 7.1 GB | 50 ms | 7.3% |
| 78 | vLLM-SR Decision 2.0 Nox 4B vLLM Semantic Router · based on Qwen3.5-4B-Base |
45.0 | -16.6 | LoRA + decision head | 4.2 B | Yes | 3.7 GB | 34 ms | 3.0% |
| 79 | frontier-infra Jebadiah 9B v2.1 Frontier Infra · based on Qwen3.5-9B |
44.9 | -16.6 | LoRA adapter | 9.7 B | Yes | 7.1 GB | 49 ms | 5.8% |
| 80 | jiwo 4B v2 · based on Qwen3.5-4B |
44.8 | -16.7 | Fine-tuned model | 4.2 B | Yes | 3.7 GB | 50 ms | 6.8% |
| 81 | Sieve 9B Sthānika AI · based on Qwen3.5-9B |
44.6 | -16.9 | LoRA + decision head | 9.7 B | Yes | 7.1 GB | 50 ms | 4.3% |
| 82 | JPT-9B · based on Qwen3.5-9B-Base |
44.5 | -17.1 | LoRA adapter | 9.7 B | Yes | 7.1 GB | 149 ms | 7.7% |
| 83 | Aplomb 1 EmpirioLabs AI · based on Qwen3.5-4B |
43.8 | -17.8 | LoRA + decision head | 5.3 B | Yes | 4.4 GB | 54 ms | 7.5% |
| 84 | FlyMy.AI Decision 4B v1.6 FlyMy.AI · based on Qwen3.5-4B |
43.6 | -17.9 | LoRA adapter | 4.7 B | Yes | 4.0 GB | 27 ms | 4.6% |
| 85 | Kev 9B v2 · based on Qwen3.5-9B-Base |
43.3 | -18.3 | LoRA + decision head | 9.7 B | Yes | 7.1 GB | 60 ms | 4.5% |
| 86 | Hopper (G) 1.2 HopitAI · based on Qwen3.5-4B-Base |
42.6 | -18.9 | LoRA adapter | 4.7 B | Yes | 4.0 GB | 23 ms | 9.3% |
| 87 | Jev-Omni · based on gemma-4-12B |
42.4 | -19.2 | LoRA + decision head | 12 B | Yes | 8.5 GB | 55 ms | 16.1% |
| 88 | djev · based on diffusiongemma-26B-A4B-it |
41.8 | -19.8 | Technique on a stock model | 25.8 B | No | 17 GB | 84 ms | 21.2% |
| 89 | JPT-4B · based on Qwen3.5-4B-Base |
41.6 | -19.9 | LoRA adapter | 4.7 B | Yes | 4.0 GB | 139 ms | 7.9% |
| 90 | JevK5 v0.3 · based on Qwen3.5-4B |
41.5 | -20.0 | LoRA adapter | 4.7 B | Yes | 4.0 GB | 26 ms | 6.0% |
| 91 | Bespoke Nimble 9B v2 Bespoke Labs · based on Qwen3.5-9B-Base |
41.0 | -20.6 | LoRA adapter | 9.7 B | Yes | 7.1 GB | 77 ms | 2.4% |
| 92 | Solomon v1.1 Doccy Health · based on Qwen3.8-27B |
40.9 | -20.7 | Decision head | 27.8 B | Yes | 18 GB | 224 ms | 8.1% |
| 93 | Decider 4B · based on Qwen3.5-4B-Base |
40.3 | -21.2 | Fine-tuned model | 4.7 B | Yes | 4.0 GB | 13 ms | 8.4% |
| 94 | Saracura PT-BR 4B Felhen · based on Qwen3.5-4B-Base |
40.2 | -21.4 | LoRA + decision head | 4.7 B | Yes | 4.0 GB | 62 ms | 17.9% |
| 95 | razorback16 openjev diffusiongemma (NVFP4, vLLM) · based on diffusiongemma-26B-A4B-it |
40.0 | -21.5 | Technique on a stock model | 25.8 B | No | 17 GB | 38 ms | 23.2% |
| 96 | Standard One 3B SH (v2) Standard Thinking · based on StandardOne-3B |
39.9 | -21.7 | LoRA + decision head | 3.8 B | Yes | 3.5 GB | 76 ms | 10.1% |
| 97 | Kev 4B r10 · based on Qwen3.5-4B-Base |
39.5 | -22.0 | LoRA + decision head | 4.7 B | Yes | 4.0 GB | 59 ms | 3.2% |
| 98 | LiquidAI d1-3B Liquid AI · based on LFM2.5-VL-3B |
38.9 | -22.6 | Fine-tuned model | 3.1 B | Yes | 3.1 GB | 16 ms | 4.7% |
| 99 | vLLM-SR d3-nano (2B) vLLM Semantic Router · based on Qwen3.5-2B |
38.9 | -22.7 | Fine-tuned model | 2.2 B | Yes | 2.5 GB | 38 ms | 11.2% |
| 100 | Winnow-E4B · based on gemma-4-E4B |
38.7 | -22.8 | LoRA adapter | 8 B | Yes | 6.0 GB | 45 ms | 5.8% |
| 101 | Lev1 3.7B · based on K2-Horizon-3.7B |
38.6 | -22.9 | LoRA adapter | 3.7 B est. | Yes | 3.4 GB | 29 ms | 5.8% |
| 102 | RSI-Jev v5.0 VL 3B · based on Qwen3.5-4B-Base |
38.4 | -23.1 | Fine-tuned model | 4.7 B | Yes | 4.0 GB | 19 ms | 4.8% |
| 103 | Intern-Decision-4B Intern Large Models · based on Qwen3.5-4B-Base |
38.2 | -23.3 | Fine-tuned model | 4.7 B | Yes | 4.0 GB | 44 ms | 2.8% |
| 104 | NeoHorse-Jev-4B TokenRhythm · based on Qwen3.5-4B-Base |
38.0 | -23.6 | Decision head | 4.7 B | Yes | 4.0 GB | 49 ms | 10.4% |
| 105 | Ryotide Qwen v0.3.1 · based on Qwen3.5-4B |
37.0 | -24.5 | Technique on a stock model | 4 B | No | 3.6 GB | 81 ms | 9.2% |
| 106 | Kas 4B · based on Qwen3-4B |
36.7 | -24.8 | LoRA adapter | 4 B | Yes | 3.6 GB | 69 ms | 9.8% |
| 107 | lev Interfaze · based on Qwen3.5-4B-Base |
36.5 | -25.0 | LoRA + decision head | 4.7 B | Yes | 4.0 GB | 71 ms | 6.4% |
| 108 | mmastrac diffusiongemma (vLLM PR 57250) · based on diffusiongemma-26B-A4B-it |
36.3 | -25.2 | Technique on a stock model | 25.8 B | No | 17 GB | 127 ms | 20.2% |
| 109 | Jevstral 8B · based on Ministral-3-8B-Base-2512 |
35.9 | -25.6 | LoRA adapter | 8 B est. | Yes | 6.0 GB | 28 ms | 4.7% |
| 110 | Intelif Qwen3-4B · based on Qwen3-4B |
34.3 | -27.3 | LoRA + decision head | 4 B | Yes | 3.6 GB | 23 ms | 7.1% |
| 111 | Tev1-4B-experimental Together · based on Qwen3.5-4B-Base |
34.1 | -27.4 | Fine-tuned model | 4.7 B | Yes | 4.0 GB | 36 ms | 10.4% |
| 112 | Jobe · based on Qwen3.5-4B-Base |
33.1 | -28.5 | Technique on a stock model | 4.7 B | No | 4.0 GB | 54 ms | 12.3% |
| 113 | Gevva E4B · based on gemma-4-E4B-it |
33.0 | -28.5 | Fine-tuned model | 8 B | Yes | 6.0 GB | 53 ms | 14.5% |
| 114 | vLLM-SR d3-lite (0.8B) vLLM Semantic Router · based on Qwen3.5-0.8B |
32.3 | -29.2 | Fine-tuned model | 0.9 B | Yes | 1.7 GB | 37 ms | 9.4% |
| 115 | openvons · based on Qwen3-4B-Instruct-2507 |
32.3 | -29.3 | Technique on a stock model | 4 B | No | 3.6 GB | 21 ms | 37.1% |
| 116 | open-jev (pngwn) · based on Qwen3.5-4B-Base |
32.1 | -29.4 | LoRA adapter | 4.7 B | Yes | 4.0 GB | 112 ms | 5.9% |
| 117 | Jev-Style 2B v3 · based on Qwen3.5-2B |
31.4 | -30.1 | Fine-tuned model | 2.3 B | Yes | 2.6 GB | 70 ms | 11.8% |
| 118 | Metask-Jev-4B Metask-AI-Lab · based on Qwen3.5-4B-Base |
31.0 | -30.6 | LoRA adapter | 4.7 B | Yes | 4.0 GB | 57 ms | 10.3% |
| 119 | Sieve 2B v2 Sthānika AI · based on Qwen3.5-2B-Base |
30.8 | -30.7 | LoRA + decision head | 2.3 B | Yes | 2.6 GB | 68 ms | 9.5% |
| 120 | Cam1 · based on Qwen3.5-4B |
30.8 | -30.8 | LoRA + decision head | 4.7 B | Yes | 4.0 GB | 249 ms | 6.3% |
| 121 | OneJev1.5-0.8B OmniJev · based on Qwen3.5-0.8B |
29.0 | -32.5 | Fine-tuned model | 1.1 B | Yes | 1.9 GB | 83 ms | 3.5% |
| 122 | SemIf · based on Qwen3.5-4B-Base |
28.8 | -32.7 | Technique on a stock model | 4.7 B | No | 4.0 GB | 113 ms | 9.9% |
| 123 | Gevva E2B · based on gemma-4-E2B-it |
28.2 | -33.4 | Fine-tuned model | 5.1 B | Yes | 4.3 GB | 43 ms | 9.0% |
| 124 | vera-core · based on Qwen3.5-2B |
28.1 | -33.5 | Fine-tuned model | 1.9 B | Yes | 2.4 GB | 100 ms | 12.2% |
| 125 | vLLM-SR Decision 2.0 Sol 2B vLLM Semantic Router · based on Qwen3.5-2B |
27.1 | -34.5 | Fine-tuned model | 2.3 B | Yes | 2.6 GB | 26 ms | 3.7% |
| 126 | Decider 2B · based on Qwen3.5-2B-Base |
25.8 | -35.7 | Fine-tuned model | 2.3 B | Yes | 2.6 GB | 8 ms | 7.7% |
| 127 | mini-jev · based on Qwen3-4B-Instruct-2507 |
25.2 | -36.4 | Technique on a stock model | 4 B | No | 3.6 GB | 67 ms | 34.6% |
| 128 | jiwo 0.8B v2 · based on Qwen3.5-0.8B |
24.9 | -36.7 | Fine-tuned model | 0.8 B | Yes | 1.7 GB | 36 ms | 6.8% |
| 129 | this-that 1.2 FLock.io · based on decider-2b |
24.6 | -37.0 | Fine-tuned model | 1.9 B | Yes | 2.4 GB | 44 ms | 19.8% |
| 130 | EXAONE-4.0-1.2B-JEV v0.3 · based on EXAONE-4.0-1.2B |
23.3 | -38.2 | Fine-tuned model | 1.3 B est. | Yes | 2.0 GB | 149 ms | 7.9% |
| 131 | Sieve 0.8B Sthānika AI · based on Qwen3.5-0.8B-Base |
23.1 | -38.5 | Fine-tuned model | 0.6 B est. | Yes | 1.6 GB | 52 ms | 8.8% |
| 132 | Intern-Decision-2B Intern Large Models · based on Qwen3.5-2B-Base |
22.6 | -39.0 | Fine-tuned model | 2.3 B | Yes | 2.6 GB | 34 ms | 6.1% |
| 133 | Crivo-150M · based on ettin-reranker-150m-v1 |
22.2 | -39.3 | Fine-tuned model | 0.1 B | Yes | 1.3 GB | 13 ms | 14.3% |
| 134 | Strands Decider 2B Strands Agents · based on Qwen3.5-2B-Base |
21.3 | -40.2 | LoRA + decision head | 1.9 B | Yes | 2.4 GB | 52 ms | 4.4% |
| 135 | Saina Helm 2.0 Saina · based on Qwen3.5-0.8B |
20.8 | -40.7 | LoRA + decision head | 0.9 B | Yes | 1.7 GB | 39 ms | 28.9% |
| 136 | Sieve 2B Sthānika AI · based on Qwen3.5-2B-Base |
20.5 | -41.0 | LoRA + decision head | 2.3 B | Yes | 2.6 GB | 63 ms | 8.8% |
| 137 | vLLM-SR d3-edge (0.6B) vLLM Semantic Router · based on Qwen3.5-0.8B |
20.2 | -41.3 | Fine-tuned model | 0.6 B | Yes | 1.6 GB | 19 ms | 14.1% |
| 138 | JPT-0.8B · based on Qwen3.5-0.8B-Base |
18.7 | -42.8 | LoRA adapter | 0.9 B | Yes | 1.7 GB | 118 ms | 6.9% |
| 139 | Sifr 0.8B v3.1 · based on Qwen3.5-0.8B |
17.8 | -43.7 | Fine-tuned model | 0.9 B | Yes | 1.7 GB | 35 ms | 12.0% |
| 140 | tasksource Decider nano v4 (150M) tasksource · based on ettin-reranker-150m-v1 |
17.6 | -43.9 | Fine-tuned model | 0.1 B | Yes | 1.3 GB | 14 ms | 6.9% |
| 141 | vLLM-SR Decision 2.0 Eos 0.8B vLLM Semantic Router · based on Qwen3.5-0.8B |
17.5 | -44.1 | Fine-tuned model | 0.9 B | Yes | 1.7 GB | 25 ms | 13.0% |
| 142 | DecisionTune 1.0 DecisionTune · based on ModernBERT-large |
16.7 | -44.8 | Fine-tuned model | 0.4 B | Yes | 1.4 GB | 14 ms | 14.8% |
| 143 | Bosun v3.1 1.7B Hanno Labs · based on Qwen3-1.7B-Base |
15.8 | -45.7 | LoRA + decision head | 1.7 B | Yes | 2.2 GB | 40 ms | 13.0% |
| 144 | Bekko System One v0 400M · based on ettin-encoder-400m |
15.2 | -46.3 | Fine-tuned model | 0.4 B est. | Yes | 1.4 GB | 17 ms | 11.1% |
| 145 | Dinah-0 · based on moBERTo |
13.9 | -47.7 | Fine-tuned model | 0.1 B | Yes | 1.3 GB | 14 ms | 10.4% |
| 146 | Kev 0.8B r15 · based on Qwen3.5-0.8B-Base |
13.7 | -47.9 | LoRA + decision head | 0.9 B | Yes | 1.7 GB | 45 ms | 4.0% |
| 147 | vLLM-SR Decision 2.0 Kai 0.6B vLLM Semantic Router · based on Qwen3-0.6B-Base |
13.6 | -47.9 | Fine-tuned model | 0.6 B | Yes | 1.6 GB | 19 ms | 3.8% |
| 148 | Tev1-0.8B-experimental Together · based on Qwen3.5-0.8B-Base |
12.0 | -49.5 | Fine-tuned model | 0.9 B | Yes | 1.7 GB | 29 ms | 12.6% |
| 149 | Jev-Style 0.8B v3 · based on Qwen3.5-0.8B |
11.9 | -49.6 | Fine-tuned model | 0.9 B | Yes | 1.7 GB | 36 ms | 8.4% |
| 150 | Crivo-32M · based on ettin-reranker-32m-v1 |
11.9 | -49.7 | Fine-tuned model | — | Yes | — | 7 ms | 8.7% |
| 151 | Kodiak v0.2 1B (accuracy) Cortex Agent LLC · based on ettin-encoder-1b |
11.8 | -49.7 | Fine-tuned model | 3 B | Yes | 3.0 GB | 142 ms | 7.7% |
| 152 | Intern-Decision-0.8B Intern Large Models · based on Qwen3.5-0.8B-Base |
11.8 | -49.8 | Fine-tuned model | 0.9 B | Yes | 1.7 GB | 34 ms | 2.5% |
| 153 | vera-spark · based on ModernBERT-large |
11.5 | -50.1 | Fine-tuned model | 0.6 B | Yes | 1.6 GB | 80 ms | 15.7% |
| 154 | GLiNER2.5-Decide Fastino · based on gliner2-large-v1 |
9.2 | -52.3 | Fine-tuned model | 0.5 B | Yes | 1.5 GB | 23 ms | 8.8% |
| 155 | LiquidAI d1-omni-600M Liquid AI · based on LFM2.5-Encoder-350M |
9.1 | -52.4 | Fine-tuned model | 0.6 B | Yes | 1.6 GB | 11 ms | 12.8% |
| 156 | Bosun v3.1 0.6B Hanno Labs · based on Qwen3-0.6B-Base |
8.7 | -52.8 | LoRA + decision head | 0.6 B | Yes | 1.6 GB | 39 ms | 14.2% |
| 157 | Bekko System One v0 68M · based on ettin-encoder-68m |
8.4 | -53.1 | Fine-tuned model | 0.1 B est. | Yes | 1.3 GB | 12 ms | 11.3% |
| 158 | Von 1.3 · based on ModernBERT-large |
8.2 | -53.4 | Fine-tuned model | 0.4 B | Yes | 1.4 GB | 18 ms | 21.6% |
| 159 | jeff · based on gliformer-large-v1 |
7.7 | -53.8 | Technique on a stock model | 0.6 B est. | No | 1.6 GB | 22 ms | 9.7% |
| 160 | JevK5 Lite · based on deberta-v3-large |
6.3 | -55.2 | Fine-tuned model | 0.3 B | Yes | 1.4 GB | 19 ms | 14.9% |
| 161 | CLM-v0.1-8B · based on Qwen3-8B-Base |
6.2 | -55.3 | Fine-tuned model | 8.2 B | Yes | 6.2 GB | 47 ms | 32.3% |
| 162 | LFM2.5-2.6B-RLCD · based on LFM2.5-2.6B-Base |
6.1 | -55.4 | Fine-tuned model | 2.7 B | Yes | 2.8 GB | 39 ms | 25.4% |
| 163 | Laya | 4.4 | -57.1 | Fine-tuned model | 0.4 B | Yes | 1.4 GB | 6 ms | 14.0% |
| 164 | Lavoir Mogan AI · based on ModernBERT-large |
4.4 | -57.1 | Fine-tuned model | 0.4 B | Yes | 1.4 GB | 20 ms | 14.9% |
| 165 | DM-JEPA 1.1 Danger Labs · based on ModernBERT-base |
4.0 | -57.5 | Fine-tuned model | 0.1 B | Yes | 1.3 GB | 63 ms | 44.5% |
| 166 | Qwen-2.5-1B-RLCD · based on Qwen2.5-1.5B |
3.9 | -57.6 | Technique on a stock model | 1.5 B | No | 2.1 GB | 44 ms | 20.5% |
| 167 | Bekko System One v0 17M · based on ettin-encoder-17m |
2.9 | -58.6 | Fine-tuned model | — | Yes | — | 7 ms | 16.1% |
| 168 | Julia 1 Supersonic Labs · based on mmBERT-small |
2.6 | -59.0 | Fine-tuned model | 0.1 B est. | Yes | 1.3 GB | 6 ms | 42.0% |
| 169 | system-one-gemma · based on gemma-3-270m |
1.3 | -60.3 | LoRA + decision head | 0.3 B | No | 1.4 GB | 32 ms | 23.9% |
| 170 | Verdict · based on gliclass-modern-base-v2.0 |
0.0 | -61.5 | Fine-tuned model | 0.2 B | Yes | 1.3 GB | 11 ms | 15.4% |
| 171 | LFM2.5-350M-RLCD | 0.0 | -61.5 | Fine-tuned model | 0.4 B | Yes | 1.4 GB | 27 ms | 56.8% |
| 172 | Nirnaya 0.4B · based on loggenix-moe-0.4B-0.2A-sft-s4 |
0.0 | -61.5 | Fine-tuned model | 0.4 B | Yes | 1.4 GB | 14 ms | 4.6% |
| 173 | Zev | 0.0 | -61.5 | Technique on a stock model | — | No | — | 1 ms | 13.8% |
Decision Index: chance-corrected score from 0 (random guessing) to 100 (perfect), mixing public benchmarks (20%), private tests of the same skills (50%) and private tasks from new domains (30%). Time per decision: median, measured by the board on one 1 x NVIDIA RTX PRO 6000. Calibration error (ECE): how far the model’s stated probabilities are from how often it is right; lower is better. 4-bit needs: our estimate for the weights at Q4_K_M plus runtime. Highlighted rows beat Jev.
Top 15 decision models
Decision Index score, axis starts at 50. Dashed line: TypeSafe’s own model, Jev 1.14 (61.53).
About decision models
What is it good for?
Routing a request to the right model, classifying tickets or emails, yes/no checks, picking the next action of an agent, scoring answers. Anywhere an LLM would write a long reply just to say “B”.
Is Jev free?
Jev itself is a closed, hosted model from TypeSafe AI. The models in this ranking are open reproductions; most have weights on Hugging Face that you can download and run for free.
Where do these scores come from?
From the Jev Decision Index, a community leaderboard on Hugging Face (Decision Index decision-index-0.3, 42 public benchmarks plus private tests), last updated 16 Oct 2026. It is unofficial and not affiliated with TypeSafe AI; neither are we. We copy its scores once a night and link every model to its page.