Skip to content
System One

Jev alternatives: hosted, open-source and local options compared

Twenty-one other decision models have a hosted API: Together AI's Tev1, which answers Choice questions only; Upstage's Solar Decide, Respan's Span-01 and Span-02, Baseten's mammoth-1 and elephant-1, Liquid AI's d1 and d1-ultra, Fastino's GLiDE, meraGPT's Decider 1, Inception's Mercury Decide, Cloudflare's Clef and Clef-omni, Perplexity's pplx-decider, Hanzo AI's Kai, Nace.AI's Drex, Celeris's Celeris Decision, Subquadratic's SubQ Snap, Microsoft's Microsoft-Decision-1 and the OpenAI Decisions API, nineteen commercial models announced between 22 September and 21 October 2026; and Kev-4B on OpenRouter, an open-source model that answers all three question types. Everything else you run yourself. Laya, Kev, Nimble, CLM, Von, Fastino's GLiNER2.5-Decide, Cloudflare's Clef and Clef-omni, Perplexity's pplx-decider, Respan's Span-02, Baseten's mammoth-1 and elephant-1 and Trong Vu's Jeffan-31b-1710 ship open weights, and Liquid AI's d1-3B and d1-omni-600M, Nace.AI's Drex 1.5, Fastino's GLiDE 1.2 and H2O.ai's non-commercial H2O-Lightning-31B ship weights under their makers' own licences. SemIf, OpenJev and a dozen smaller clones read decisions off frozen open models. For fixed labels, a fine-tuned encoder still beats all of them. Independent tests show most open options trail Jev by a wide margin. On Jev Decision Index 0.3, pplx-decider-v1.1 scores above Jev, and Fastino's own benchmark puts GLiNER2.5-Decide ahead of what it calls JevK5.

Updated

The short answer

Jev has no drop-in replacement. What you pick depends on which part of Jev you need.

If you want someone else to run it, there are twenty-one other hosted decision models. Tev1 on Together AI and Kev-4B on OpenRouter cost the same $0.042 per million input tokens as Jev. Tev1 only answers Choice questions. Kev-4B answers Choice, Score and Noul and takes the same request as Jev.

The nineteen newest are commercial models. Solar Decide from Upstage answers all three question types on the same /v1/systemone request schema as Jev. It is in beta at $0.10 per million input tokens, or $0.05 on OpenRouter for a limited time. Span-01 from Respan scores agent traces for behaviors you define, at $0.02 per million input tokens, and has a free Lite tier. d1 from Liquid AI answers all three types too. Its current hosted version, d1.1, costs $0.02 per million input tokens on Liquid’s API and OpenRouter. Liquid’s larger d1-ultra, released on 15 October 2026, answers all three at $0.042 per million input tokens, the same as Jev, on Liquid’s API and OpenRouter. GLiDE from Fastino answers all three on Fastino’s /v1/systemone endpoint at $0.04 per million input tokens, and version 1.2 also reads images. Decider 1 from meraGPT also answers all three on the same schema, at $0.04 per million input tokens, and since version 1.1 on 9 November 2026 it reads images and takes 32,768 tokens for the state plus each question. Mercury Decide from Inception answers all three at $0.02 per million input tokens on Inception’s API. OpenRouter showed no endpoint for it on 12 November 2026. Inception’s own API takes POST /v1/decisions. Clef from Cloudflare answers all three on the same request schema and also reads images. It runs on Workers AI and OpenRouter at $0.24 per million input tokens, or $0.038 for the smaller Clef-flash. Cloudflare’s third model, Clef-omni, launched on 9 October 2026 at $0.15 and also reads audio and video. pplx-decider from Perplexity launched the same day and answers Choice, Score and Noul questions about text or images. It costs $0.02 per million input tokens, but it uses Perplexity’s own /v1/decisions request format instead of /v1/systemone. Kai from Hanzo AI answers Choice, Score and Noul questions at $0.021 per million input tokens, on Hanzo’s own /v1/decisions endpoint. Celeris Decision from Celeris answers all three, reads images, and takes Jev’s /v1/systemone request. Its current version, Celeris-1.1 Decision, costs $0.004 per million input tokens. Drex from Nace.AI answers all three on the /v1/systemone request at $0.03 per million input tokens for Drex Max v1.1, Drex Max Finance and Drex 2.0, and the Drex Max models read images. Microsoft-Decision-1 from Microsoft answers yes/no, multiple-choice and rating questions on Microsoft Foundry and OpenRouter at $0.042 per million input tokens, the same as Jev. Microsoft Learn publishes its request: Jev’s state and questions fields at /providers/microsoft/v1/systemone on your Foundry resource. The OpenAI Decisions API has been generally available since 22 October 2026, at $0.10 per million input tokens, with its own request format. Respan’s second model, Span-02, released on 15 October 2026, answers all three types plus number questions about text and images at $0.01 per million input tokens, on Respan’s /api/v1/decisions endpoint in the System One format. Baseten’s mammoth-1 and elephant-1, released on 21 October 2026, answer all three on Baseten’s /v1/systemone endpoint and OpenRouter. After a free first week they cost $0.10 and $0.035 per million input tokens. Subquadratic’s SubQ Snap, first released on 12 October 2026, answers all three on its own /v1/systemone endpoint at $0.04 per million input tokens and reads up to 384K tokens in one request. It is in public preview. The remaining hosted route is an ordinary LLM with structured outputs, which is slower and costs more per decision.

If you want weights you can run and keep, the serious candidates are Laya, Kev, Bespoke Nimble, CLM, Von and Fastino Labs’ GLiNER2.5-Decide, released 24 September 2026. Each one ships trained weights under Apache 2.0, as do Cloudflare’s Clef, Clef-flash and Clef-omni, which are also hosted on Workers AI, and Perplexity’s pplx-decider, which is also hosted on Perplexity’s API. Liquid AI released small open-weight d1 models, d1-3B and d1-omni-600M on 7 October 2026 and d1-2B and d1-Omni-2B on 24 October, under its LFM Open License, which limits commercial use to companies under $10 million in annual revenue. Nace.AI’s Drex 1.5 weights are on Hugging Face under Nace’s own licence, which bars use beyond personal or research work by organisations with more than $1 million in revenue or funding. Fastino put the GLiDE 1.2 weights on Hugging Face on 22 October 2026 under its Fastino Community License, which requires a separate agreement for production commercial use by companies with more than $10 million in annual revenue. Respan’s Span-02 weights are Apache 2.0 under Google’s Gemma 4 terms, and Baseten’s mammoth-1 and elephant-1 weights are MIT; all three are also hosted. Two more open-weight models rank high on the Jev Decision Index 0.4.1: Jeffan-31b-1710, an Apache 2.0 Gemma-4-31B-it fine-tune from one developer, and H2O.ai’s H2O-Lightning-31B, whose licence allows non-commercial, non-production use only. Behind them sit community clones such as SemIf, OpenJev and NanoJev. Those read decisions off open models, and most are research code a week old.

If your labels are fixed and you have training data, skip the whole category. A fine-tuned encoder classifier beat every open System One option in the one public benchmark that included one.

What “alternative” means here

People searching for a Jev alternative usually want one of three things:

  1. The same API somewhere else: you have code that calls POST /v1/systemone and you want it off TypeSafe. Many open projects speak that wire format, and so do the hosted Solar Decide, d1, Decider 1, Celeris Decision and Drex, so the change is one base URL. Clef and Clef-omni take the same fields but at a Workers AI URL, with an extra model field. Perplexity’s pplx-decider does not: it uses its own /v1/decisions format. Inception’s API also serves Mercury Decide at its own /v1/decisions endpoint, and Hanzo’s Kai uses Hanzo’s /v1/decisions. Microsoft-Decision-1 takes the same fields at /providers/microsoft/v1/systemone on a Foundry resource, with your deployment name as model.
  2. A model you own: weights on your disk, no text leaving your network, no rate limit. That is the “jev local” and “jev open source” question.
  3. The same job done another way: a classification model, an NLI model or an LLM with a JSON schema. None of these is a System One model, and all of them can sort text into labels.

No alternative copies Jev’s training. TypeSafe has not published RLCD, its method for training calibrated probabilities, so every project here copies the interface and not the method. RLCD explained covers what is known.

Jev alternatives compared

Stars were read from the GitHub API on 30 September 2026 and change daily. “Own” in the evidence column means the authors measured it themselves.

Option How you use it Base model and size Licence Stars Eval evidence
Jev 1.14 TypeSafe API, $0.042 per million input tokens Not published Closed n/a TypeSafe’s own figures; outside tests below
Tev1-4B-experimental Together AI serverless API, $0.042 per million input tokens Qwen3.5-4B, LoRA fine-tune Code MIT; the weights card declares no licence 173 880 of 1,000 on a reused dev set (own)
Solar Decide Upstage API in beta at $0.10 per million input tokens, or OpenRouter at $0.05 for a limited time Solar Mini 4, 35B mixture-of-experts with 3B active No downloadable weights n/a 95.8% on a 48-question pool against Jev’s 93.8%, from a third-party pull request; no Upstage figures
Span-01 OpenRouter, $0.02 per million input tokens; Span-01 Lite is free. Respan says Span-02 replaced it Not listed Not listed n/a 0.806 F1 against Jev’s 0.716 on Respan’s own behavior benchmark (own); 85.4% against Jev’s 93.8% on the third-party 48-question pool
d1 Liquid AI API as d1.1, $0.02 per million input tokens; OpenRouter at the same price; d1-2B, d1-Omni-2B, d1-omni-600M and the deprecated d1-3B self-hosted Not published for the hosted d1.1; LFM3-VL-2B, 1.92B (d1-2B); LFM3-Omni-2B, 2.17B (d1-Omni-2B); LFM2.5-VL-3B, 3.1B (d1-3B); LFM2.5-Encoder-350M, 587M (d1-omni-600M) Hosted d1.1 has no weights; the open models are under the LFM Open License v1.0 n/a Hosted d1: 58.9 against Jev’s 57.9 on Liquid’s own reproduction of Jev Decision Index 0.2.1 (own), not on the public index. d1-3B: 40.08 against Jev’s 60.11 on Jev Decision Index 0.3, with Liquid’s own run as the public part
d1-ultra Liquid AI API at /decisions/v1/systemone, $0.042 per million input tokens, output free; OpenRouter and Vercel’s AI Gateway at the same price; reads images Not published; Liquid says a larger open-weight model than d1 No downloadable weights n/a 70.6 against Jev 1.14’s 66.2 on the Artificial Analysis Decision Index v2.1, 5th of 22 decision models; 76.7 against Jev 1.14’s 69.4 on a 112-evaluation suite Liquid built and has not published (own); not on the Jev Decision Index
GLiDE Fastino API at /v1/systemone, $0.04 per million input tokens, output free; up to 65,536 tokens per prompt; reads images; weights also self-hostable Qwen3.8-27B, full fine-tune; the Jev Decision Index lists 27.8B Fastino Community License: production commercial use by companies over $10M in annual revenue needs a separate agreement n/a GLiDE 27B: 73.37 against Jev 1.14’s 60.24 on the Jev Decision Index 0.4.1 Full score, 2nd of 301 rows and tied for first (GLiDE 1.2 led the 22 October file at 71.52); 72.5 against Jev 1.14’s 66.2 on the Artificial Analysis Decision Index v2.1, 3rd of 22; 84.5 Capability Score on JevBench for the hosted API (22 October). Launch version: 64.81 against Jev’s 57.91 on Jev Decision Index 0.2.1, run by Fastino (own); 53.9 against Jev’s 57.5 on the Artificial Analysis Decision Index v1.0
Decider 1 meraGPT API at $0.04 per million input tokens; version 1.1 takes 32,768 tokens for the state plus each question and reads images Not published No downloadable weights n/a 0.768 accuracy against Jev’s 0.727 on typed-decisions, a benchmark meraGPT built (own); no outside test
Mercury Decide Inception API at /v1/decisions, $0.02 per million input tokens; no OpenRouter endpoint on 12 November 2026 Not published No downloadable weights n/a Inception calls it the top decision model on OpenRouter by JevBench v1.4 (own, no evaluation linked); not on the Jev Decision Index
Clef Cloudflare Workers AI at $0.24 per million input tokens (Clef) or $0.038 (Clef-flash), the same through Cloudflare or PrimeIntellect on OpenRouter, or $0.192 and $0.019 through DeepInfra there; weights also self-hostable Qwen3.8-27B (Clef); Qwen3.5-9B (Clef-flash) Apache 2.0 n/a Above Jev on 8 of 10 Jev Decision Index tasks and 3 of 4 TypeSafe workflow evals, below it on the rest (own, Cloudflare’s run); Clef 53.08 and Clef-flash 47.61 against Jev’s 60.11 on Jev Decision Index 0.3
pplx-decider Perplexity Decisions API at $0.02 per million input tokens for the 27B v1.3, $0.01 for the 4B Flash v1.3; weights also self-hostable Qwen3.8-27B (Flash: Qwen3.5-4B) Apache 2.0 n/a v1: 85.71% against Jev’s 84.51% across 11 benchmarks, measured by Perplexity (own). v1.3: Jev Decision Index 0.4.1 Full score 67.67 against Jev 1.14’s 60.24; v1.2 scores 67.98, public run submitted by Perplexity
Kai Hanzo API at /v1/decisions, $0.021 per million input tokens; up to 1,024 tokens per question plus state mmBERT-base encoder initialised from Laya, per Hanzo No downloadable weights n/a 0.857 against Jev’s 0.780 mean accuracy across 12 suites, run by Hanzo on suites it trained on (own); not on the Jev Decision Index
Celeris Decision Celeris API at /v1/systemone, $0.004 per million input tokens for Celeris-1.1 Decision; reads images Not published No downloadable weights n/a Celeris-1.1 Decision: 0.802 against Jev 1.14.0’s 0.726 accuracy on jev-bench, run by Celeris (own); not on the Jev Decision Index
Microsoft-Decision-1 Microsoft Foundry (Preview), $0.042 per million input tokens; up to 32,768 input tokens, text only Qwen3.5-9B, post-trained by Microsoft No downloadable weights n/a 83.5% average accuracy over 36 benchmarks against 81.9% for Quyet-1.0-Large, run by Microsoft (own), with no accuracy score for Jev; not on the Jev Decision Index
Clef-omni Cloudflare Workers AI or OpenRouter at $0.15 per million input tokens; reads audio, video and images; weights also self-hostable Qwen3-Omni-30B-A3B, mixture-of-experts with 3B active Apache 2.0 n/a Above Jev on 8 of 10 Jev Decision Index tasks, below it on 4 of 5 TypeSafe workflow evals and tied on one (own, Cloudflare’s run); Jev Decision Index Full score 49.32 against Jev’s 60.11 (Index lab run)
Drex Nace.AI API at /v1/systemone, $0.03 per million input tokens for Drex Max v1.1, Drex Max Finance and Drex 2.0; OpenRouter through DeepInfra; the Drex Max models read images Drex 2.0: Qwen3.6-35B-A3B, 35B with about 3B active; Drex Max: weights and base not published by Nace Nace.AI Open RAIL-M for the 2.0 and 1.5 weights, commercial use limited to small organisations 3 Drex Max 67.6 on the public Jev Decision Index 0.3.2 suite (own); Jev Decision Index 0.4.1 Full score 66.6 for Drex Max and 52.76 for Drex 2.0 against Jev 1.14’s 60.24, public runs submitted by Nace
OpenAI Decisions API OpenAI API at /v1/decisions, generally available, or OpenRouter, $0.10 per million input tokens; reads images GPT-6 Luna No downloadable weights n/a Right control in 76 of 78 steps against Jev’s 73 on Every’s preview test (testers’ own); 60.2 against Jev 1.14’s 66.2 on the Artificial Analysis Decision Index v2.1; not on the Jev Decision Index
Span-02 Respan API at /api/v1/decisions or OpenRouter, $0.01 per million input tokens, output free; reads images; weights also self-hostable on a Blackwell GPU Gemma 4 26B-A4B, mixture-of-experts with about 4B active Apache 2.0, weights also under the Gemma 4 terms n/a 59.07 against Jev 1.14’s 60.24 on the Jev Decision Index 0.4.1 Full score, public run submitted by the author; 69.3 Capability Score on JevBench
mammoth-1 Baseten Model APIs at /v1/systemone or OpenRouter, $0.10 per million input tokens since a free week that ended 27 October 2026; weights also self-hostable Qwen3.5-27B, full fine-tune, 27.4B MIT n/a 68.44 against Jev 1.14’s 60.24 on the Jev Decision Index 0.4.1 Full score, 22nd of 301, public run submitted by Base Labs; 85.2 Capability Score on JevBench’s open board; 71.4 against Jev 1.14’s 66.2 on the Artificial Analysis Decision Index v2.1, 4th of 22; 65.9 on Base Labs’ own replica of the Artificial Analysis index (own)
elephant-1 Baseten Model APIs at /v1/systemone or OpenRouter, $0.035 per million input tokens since a free week that ended 27 October 2026; weights also self-hostable Qwen3.5-9B, full fine-tune, 9.4B MIT n/a 66.4 against Jev 1.14’s 66.2 on the Artificial Analysis Decision Index v2.1, 10th of 22; 60.26 against 60.24 on the Jev Decision Index 0.4.1 Full score, trained on Index train splits by the authors’ own account
SubQ Snap SubQ API at /v1/systemone, $0.04 per million input tokens, output free; current version 1.2; reads images; 384K-token context Not published No downloadable weights n/a SubQ Snap 1.2: 72.6 against Jev 1.14’s 66.2 on the Artificial Analysis Decision Index v2.1, 2nd of 22 (1.1: 67.9); JevBench Capability Score 81.5 for the first release; 65.7 against Jev 1.14’s 60.6 on the Jev Decision Index 0.2.1 suite (own)
Laya Self-hosted, GPU or CPU ModernBERT-large, 421M; mmBERT-base, 322M Apache 2.0 28,759 TREC 86.6%, BANKING77 37.0% (SOTAAZ); Berkeley exam 31.2%
Kev Self-hosted, CUDA, ROCm or Apple MLX; Kev-4B also on OpenRouter at $0.042 per million input tokens Qwen3.5 at 0.8B, 4B and 9B; Qwen3.8-27B for the 27B Apache 2.0 7,948 Kev-27B 0.848 and Kev-9B 0.822 against Jev’s 0.857 on unseen datasets (own)
Jeffan-31b-1710 Self-hosted, one 96 GB GPU, vLLM Gemma-4-31B-it, full fine-tune, 31.3B Apache 2.0 (CC BY-NC 4.0 until 22 October 2026) n/a 69.75 against Jev 1.14’s 60.24 on the Jev Decision Index 0.4.1 Full score, 17th of 301, public run submitted by the author; 81.7 Capability Score on JevBench; training data undisclosed
H2O-Lightning-31B Self-hosted, one RTX PRO 6000 or H100 80GB, vLLM Gemma-4-31B-it, full fine-tune, 31.3B H2O Model Non-Commercial Licence v1.1, non-commercial and non-production use only n/a 69.92 for v1.3.0 against Jev 1.14’s 60.24 on the Jev Decision Index 0.4.1 Full score, 16th of 301; 84.3 Capability Score on JevBench’s open board for v1.0.7
GLiNER2.5-Decide Self-hosted, CPU or GPU; Fastino also hosts inference at agent.fastino.ai DeBERTa-v3-large fine-tune, 340M per the model card (Hugging Face’s own weight metadata reports 486M) Apache 2.0 2,237 60.1% on Fastino’s own 17-dataset suite, ahead of JevK5 at 57.5%
Bespoke Nimble Self-hosted, Apple Silicon or NVIDIA Qwen3.5-9B, LoRA fine-tune Apache 2.0 on the model card 1,915 90.12% against Jev’s 93.21% on 324 held-out examples (own)
CLM Self-hosted, Linux and NVIDIA Qwen3-8B plus a 75MB head Apache 2.0 2,540 Claims parity with Jev at up to 9x the speed (own, unreproduced)
Von Self-hosted, GPU or CPU ModernBERT, 395M Apache 2.0 776 Von 1.1 72.0% against Jev’s 96.6% on a 49-task suite (own)
Decider Self-hosted, GPU Qwen3.5-2B-Base, fine-tuned Apache 2.0 979 None published
SemIf (formerly openjev) Self-hosted, RTX 3090, Mac or CPU Frozen Qwen3.5-4B MIT code 4,583 Berkeley exam 61.6%; 0.845 agreement on a TypeSafe subset where Jev gets 0.883 (own)
OpenJev (razorback16) Self-hosted, or free on Codiv Frozen DiffusionGemma 26B-A4B Apache 2.0 540 BANKING77 66.9%, TREC 66.0%, AG News 85.0% (SOTAAZ)
NanoJev Self-hosted, CUDA Qwen3-0.6B MIT 2,440 TREC 18.4%, BANKING77 20.1% (SOTAAZ); trained on games
jevlike Self-hosted, CPU or GPU Own small byte-level model MIT 1,336 None published
Rizzo Flow Self-hosted through llama.cpp Frozen Spark-X2.5, 4B or 1.7B Apache 2.0 758 Tied with SemIf on SemIf’s own test set (own)

Sources: each project’s README or model card, the model pages on this site, and the SOTAAZ and Berkeley exam write-ups linked below, all checked 24 September 2026, except that star counts, licences and the Von and Kev rows were rechecked on 30 September 2026. The Solar Decide, Span-01 and Decider 1 rows come from their model pages, checked 30 September 2026. The d1, Mercury Decide, Clef and pplx-decider rows come from their model pages and Jev Decision Index 0.3, and the Kai and Celeris Decision rows from their model pages, all checked 9 October 2026. The Microsoft-Decision-1, Clef-omni and Drex rows come from their model pages, checked 10 October 2026; Drex’s star count is for nace-ai/drex-decision-models on that date. The OpenAI Decisions API and d1-ultra rows come from their model pages, last updated 14 to 16 October 2026, and the GLiDE row from its model page, updated 23 October 2026. The Span-02, mammoth-1, elephant-1, Jeffan-31b-1710 and H2O-Lightning-31B rows come from their model pages, checked 23 October 2026. The pplx-decider row was rechecked against its model page and Jev Decision Index 0.4.1 on 14 November 2026.

The Jev Decision Index

The broadest independent test so far is the Jev Decision Index (multimodalart’s community board on Hugging Face) by multimodalart. Its 0.3.1 edition, generated on 10 October 2026, sends 109,759 scored questions to Jev and 116 other entries. The text scores are unchanged from edition 0.3 of 7 October. It runs open entries on a single RTX PRO 6000 and publishes its code.

Edition 0.3 changed what the headline score means, so its numbers cannot be set against those of 0.2.1. In 0.2.1 the headline was the public benchmarks alone, chance-corrected so that 0 is random guessing and 100 is perfect, over 38 benchmarks. In 0.3 the headline is a Full score: 20% from 37 public benchmarks, 50% from private tests of the same skills and 30% from private tasks in new domains, with the private parts mapped onto the public scale. The public part is still published on its own as the Public only score, but its suite changed too: GSM8K was rebuilt, ForecastBench retired and WinoGrande moved to another area. Entries within 0.9 points share a rank. Calibration error is measured as before, on 32 benchmarks, and entries that did not change keep the same figure.

Entry Full score Public only Calibration error
pplx-decider-v1.1-27b 62.75 62.25 0.071
Jev 1.13 60.11 57.96 0.074
Kev 27B 58.78 56.69 0.022
Decider chat with Gemma-4-31B 57.62 57.79 0.047
Surogate Rune 26B-A4B v3 57.43 58.28 0.120
Bespoke Nimble 9B v3 54.67 57.19 0.123
Clef 53.08 61.71 0.040
Decider 35B-A3B 48.53 47.70 0.023
Clef-flash 47.61 56.15 0.025
Kev 9B v2 43.27 41.54 0.045
d1-3B 40.08 48.99 0.047
Tev1-4B-experimental 34.14 29.51 0.104
d1-omni-600M 9.45 17.90 0.128
GLiNER2.5-Decide 9.24 11.48 0.088
CLM-v0.1-8B 6.24 7.41 0.323
Laya 4.43 6.19 0.140

Calibration error is the average gap between a model’s stated confidence and how often it is right. Perplexity’s pplx-decider-v1.1-27b leads the Full score on its own. Jev shares second place with two entries that are not in this table, both within 0.9 points of it. Kev 27B is new on the board. The private tests move some entries a long way: Clef is third of 117 on Public only but 53.08 on the Full score, and Clef-flash falls from 56.15 to 47.61. pplx-decider v1 left the board when v1.1 was scored. v1.1’s calibration error is 0.071, where v1 had 0.018 in 0.2.1.

Liquid AI submitted the public runs for d1-3B and d1-omni-600M itself, on an AMD MI325X, and the Index ran the private tests. The hosted d1, Mercury Decide, Solar Decide, Span-01, meraGPT’s Decider 1, Kai, Celeris Decision, Microsoft-Decision-1 and the OpenAI Decisions API are not on the board. Drex and Clef-omni have rows. Nace.AI submitted the public run for Drex, which scores 50.84 on the Full score and 58.08 on Public only, and Nace says it trained on the benchmarks’ training splits. Clef-omni is a lab run by the Index, with no submission from Cloudflare, at 49.32 and 56.54. The two Decider rows are the unrelated open Decider. The small encoders that do well on narrow label sets score near random on this broad suite. It is one person’s suite and it changes between editions, so check the current page.

Hosted options

Tev1 on Together AI

Tev1 is a System One style model you can call as a paid service on Together AI. Together fine-tuned Qwen3.5-4B to read a state, a question and 2 to 24 options, then return one option letter. It costs $0.042 per million input tokens with free output, the same as Jev.

It is a narrower product. It has no Score and no Noul. The model card says it “retains Qwen’s standard next-token language-model head,” so it is a normal language model trained to answer with one letter, not a model that skips text generation. The repo says its logprobs are not calibrated confidence. Its one published result, 880 of 1,000, comes from a development set the team reused while building the recipe. Together also published the recipe and says a training run costs about $17. That makes Tev1 as much a template for training your own classifier as a product.

Solar Decide on Upstage

Solar Decide is Upstage’s System One model, in beta since 22 September 2026. It runs on Solar Mini 4, a 35B mixture-of-experts model with 3B parameters active per token, and reads a state of up to 512K tokens. It answers Choice, Score and Noul questions and uses the same /v1/systemone request schema as Jev. Upstage lists $0.10 per million input tokens with free output. OpenRouter added it on 28 September at $0.05, a discount its launch post calls limited-time.

A Choice takes 2 to 26 options, and Upstage warns that the schema may change before general availability. Upstage publishes no accuracy figures. The only measurements are third-party. In a pull request to zero-shot-ie-bench, its author reports 95.8% on a 48-question sentiment and topic pool against 93.8% for Jev. That is a small sample from one person’s test set.

Span-01 on Respan

Span-01 does a different job from Jev. You send a conversation span and behaviors written in plain language, such as “the user expresses frustration,” and it returns the probability that each is present, absent or not observable. Noul is the closest System One type, but the output does not match it exactly.

Respan’s launch post now says Span-02 replaced Span-01, and its Span-01 docs redirect to Span-02. OpenRouter still listed Span-01 at $0.02 per million input tokens with free output, and Span-01 Lite for free, on 14 November 2026. On Respan’s own behavior benchmark Span-01 scores 0.806 F1 against 0.716 for Jev. That is Respan’s number on its own dataset. The one outside test, on a 48-question sentiment and topic pool, has Span-01 at 85.4% and Lite at 79.2% against 93.8% for Jev.

d1 from Liquid AI

d1 is Liquid AI’s first decision model, announced on 29 September 2026. It answers Choice, Score and Noul questions and writes no text. Liquid’s docs call POST https://api.liquid.ai/decisions/v1/systemone with TypeSafe’s Python and TypeScript SDKs, using the model id d1.1, which reads text and images. At launch the ids were d1:free and the paid d1. Liquid’s pricing page lists d1.1 at $0.02 per million input tokens, and Liquid has not published the base model or the parameter count.

Its launch post says d1 beats Jev on multilingual evals, prompt injection resistance and longer inputs, and gives no numbers for any of the three. The one chart is Liquid’s own reproduction of the Jev Decision Index, where d1 leads Jev overall but trails it on Knowledge, 43.3 against 51.3. Nobody outside Liquid has reproduced it, and the hosted d1 is not on Jev Decision Index 0.3. OpenRouter has listed it as liquid/d1 since 1 October 2026, at first for $0.04 per million input tokens and now $0.02, and added liquid/d1.1 on 27 October at $0.02, with free output. Vercel’s AI Gateway lists it under the same id.

On 7 October 2026 Liquid released two open-weight models in the d1 family. d1-3B is a 3.1B model on LFM2.5-VL-3B that reads text and images. d1-omni-600M is a 587M model on the LFM2.5-Encoder-350M encoder that reads text with images or with up to 30 seconds of audio. Both are on Hugging Face under the LFM Open License v1.0, and Liquid’s docs do not offer either on its API. On Jev Decision Index 0.3, d1-3B scores 40.08 and d1-omni-600M 9.45, against Jev’s 60.11. Liquid ran the public part of both itself. On 24 October Liquid added d1-2B, a 1.92B model on LFM3-VL-2B that also reads video, and d1-Omni-2B, a 2.17B model on LFM3-Omni-2B that reads video and audio. Liquid’s docs now mark d1-3B deprecated in favour of d1-2B.

Decider 1 from meraGPT

Decider 1 is meraGPT’s decision model, announced on 22 September 2026. It answers Choice, Score and Noul questions and writes no text. It serves the same /v1/systemone request schema as Jev at https://meragpt.com, so TypeSafe’s SDKs work once you change the base URL. It costs $0.04 per million input tokens with free output; the launch post gave $0.03. Decider 1.1, released on 9 November 2026, now answers every request. A Choice takes up to 255 labels, a request up to 64 questions, and the state plus each question can hold 32,768 tokens, up from 4,096 for the whole request. The state can also hold images. It is a hosted API only, with no weights. It is not related to the open Decider covered below.

Every accuracy figure comes from meraGPT’s own run on typed-decisions, a synthetic benchmark that meraGPT told this site it built and that it published under the LocalLLaMA organisation on Hugging Face. There Decider 1 scores 0.768 against Jev’s 0.727. The benchmark’s own card says scores well above 0.735 mean a model is learning the quirks of the teacher model that wrote the answers. Nobody outside meraGPT had measured Decider 1 by 30 September 2026, and it is not on Benchmark Heaven’s JevBench.

Mercury Decide from Inception

Mercury Decide is the decision model from Inception, which makes the Mercury diffusion language models. It launched on 30 September 2026 and answers Choice, Score and Noul questions with a probability, writing no text. Inception’s own API takes POST https://api.inceptionlabs.ai/v1/decisions with the model mercury-decide, and its docs say it does not support chat completions. A Choice takes 1 to 255 options and a Score 1 to 10 levels.

It costs $0.02 per million input tokens with free output on Inception’s API, which Inception lists as half its regular $0.04. OpenRouter sold it at the same price as inception/mercury-decide, with a free inception/mercury-decide:free, but showed no endpoint for either on 12 November 2026. Inception calls it the most intelligent decision model on OpenRouter by JevBench v1.4, without linking the evaluation, and it is not on the Jev Decision Index. There are no published weights.

Clef from Cloudflare

Clef is Cloudflare’s own decision model, launched on 1 October 2026 in two sizes: Clef at 27B on Qwen3.8-27B and Clef-flash at 9B on Qwen3.5-9B. Both run on Workers AI and take Jev’s state and questions request with Noul, Choice and Score questions. Clef adds an optional images array. On Workers AI, Clef has a 65,536-token context and costs $0.24 per million input tokens; Clef-flash has 24,576 tokens and costs $0.038, checked on 10 October 2026. OpenRouter serves both through Cloudflare and PrimeIntellect at the same prices, and through DeepInfra in fp8 at $0.192 and $0.019, checked on 12 November 2026. The weights are on Hugging Face under Apache 2.0.

Cloudflare’s launch numbers are its own run. On 10 tasks it chose from the Jev Decision Index, Clef beats Jev on eight and loses on When2Call (72.37 against 80.97) and BRIGHT (45.91 against 47.52). Clef-flash scores 66.77 on CLINC150+OOS against Jev’s 89.27. On TypeSafe’s workflow evals Clef wins three of four and loses agent trace observability, 68.5 to 71.6. Cloudflare reports a median latency of 209.3ms for Clef and 38.8ms for Clef-flash, against 524.1ms for Jev, in its 1 October post. On 9 October it reported a faster hosted Clef, with a median of 152ms at about 800 tokens, down from 262ms, and no new Jev figure. It calls its second training stage RLCD, which it describes itself and does not say matches TypeSafe’s method.

The first outside run is the Jev Decision Index (0.3.1 file, 10 October 2026). On its public benchmarks Clef scores 61.71, ahead of Jev’s 57.96, and Clef-flash 56.15. On the Full score, which adds private tests, Clef falls to 53.08 and Clef-flash to 47.61, against Jev’s 60.11. Their calibration errors are 0.040 and 0.025, against Jev’s 0.074.

Clef-omni, launched on 9 October 2026, is Cloudflare’s third model. It is post-trained from Qwen3-Omni-30B-A3B, a mixture-of-experts model with 3B parameters active, and adds optional audio and videos arrays to the same request. It costs $0.15 per million input tokens on Workers AI and OpenRouter, with a 64,000-token context on Workers AI. In Cloudflare’s own tables it beats Jev on eight of 10 Jev Decision Index tasks but trails it on four of five workflow evals and ties on the fifth. The Jev Decision Index ran it in a lab run and scores it 49.32 on the Full score, 27th of 117 rows against 60.11 for Jev, and 56.54 on Public only, against Jev’s 57.96.

Cloudflare’s Clairic, in preview since 19 October 2026, is not a Jev alternative. It is an extraction model, a different System One class that reads values out of the input, and it pairs with a decision model rather than replacing one.

pplx-decider from Perplexity

pplx-decider is Perplexity’s decision model. The first version, pplx-decider-v1-27b, launched on 1 October 2026, the same day as Clef, on a Qwen3.8-27B base with Apache 2.0 weights and image input. Since 12 November 2026 the current versions are the 27B pplx-decider-v1.3 and the 4B pplx-decider-flash-v1.3, both on Hugging Face, and v1.2 is still served by name. That is where the overlap ends. The Decisions API takes named questions at POST /v1/decisions, not Jev’s /v1/systemone, with no general cap per call as of 14 November 2026. It answers Noul, Choice and Score questions, and on v1.3 locate questions that return spans of the text. The state and each question can hold up to 262,144 input tokens, four times Clef’s limit, and the rate limit is 10 requests per second per organization.

It costs $0.02 per million input tokens for the 27B model and $0.01 for Flash as of 14 November 2026, with output free, down from $0.04 at launch. The v1 model card reports 85.71% across 11 benchmarks against 84.51% for Jev, with Jev ahead on 6 of the 11. Perplexity ran these itself and does not say how it ran Jev, so they are vendor-run. On the independent Jev Decision Index 0.2.1, checked 2 October 2026, v1 scored 56.40 against Jev’s 57.91, with a calibration error of 0.018 against Jev’s 0.074. The v1.1 card reports 61.56 against Jev’s 57.9 on the 0.2.1 suite, from Perplexity’s own run. Jev Decision Index 0.3 puts v1.1 first on its Full score, at 62.75 against Jev’s 60.11, with a calibration error of 0.071. That score uses a new formula, so it does not compare with the 0.2.1 figures. In edition 0.4.1, checked on 14 November 2026, v1.3 scores 67.67 on the Full score and v1.2 67.98, against Jev 1.14’s 60.24. Perplexity’s Python and TypeScript SDKs now have a decisions.create method.

Kai from Hanzo AI

Kai is Hanzo AI’s decision model, launched on 28 September 2026. It answers Choice, Score and Noul questions and writes no text. The request is POST https://api.hanzo.ai/v1/decisions with the model set to kai, not Jev’s /v1/systemone. A request holds up to 100 questions, and a question plus its state can take 1,024 tokens at most. The weights are closed.

It costs $0.021 per million input tokens with free output, half Jev’s $0.042. Hanzo’s own 12-suite harness gives Kai 0.857 mean accuracy against 0.780 for Jev 1.13, but Kai trained on those suites’ train splits, so treat it as vendor-run. Hanzo says Kai is weaker on wide choices, at 0.9075 for 150 options against Jev’s 0.940. It is on Benchmark Heaven’s JevBench and the Artificial Analysis Decision Index, but not the Jev Decision Index.

Celeris Decision from Celeris

Celeris Decision is a decision model from Celeris, a lab that builds diffusion language models. Its first version, Celeris-1 Decision, launched on 8 October 2026. It answers Noul, Choice and Score questions about text, structured data or images, takes up to 8 images per request, and can add a one-sentence reason to each answer. It uses Jev’s /v1/systemone request at its own base URL, so TypeSafe’s Python SDK works once you change base_url. A second /v1/decisions endpoint follows the OpenAI format. The weights are closed.

Celeris-1.1 Decision, released on 5 November 2026, costs $0.004 per million input tokens with free output, down from $0.04. Celeris reports a median of 36 ms for one question, and 0.802 accuracy against 0.726 for Jev 1.14.0 on jev-bench, both from its own tests, so both are vendor-run. The docs give a 131,072-token context window. Benchmark Heaven’s JevBench gave the first version a Capability Score of 73.6, and it is not on the Jev Decision Index.

Drex from Nace.AI

Drex is the decision model family from Nace.AI, a Palo Alto company that sells small models to finance and audit teams. Drex launched on 24 September 2026, version 1.5 followed on 28 September, and Drex 2.0 on 6 October. Drex Max, Nace’s most accurate model, followed in mid-October, and Drex Max v1.1, now the default, on 28 October. Drex Max Finance followed on 1 November. Only the Drex Max models read images. All answer Noul, Choice and Score questions on Jev’s /v1/systemone request at https://console.nace.ai, so TypeSafe’s SDK works with a changed base URL and model name. Every active model takes up to 512 questions. Drex 2.0 is a mixture of experts post-trained from Qwen3.6-35B-A3B, and its weights are on Hugging Face under a licence that limits commercial use to small organisations. Nace has published no Drex Max weights.

Drex Max v1.1, Drex Max Finance and Drex 2.0 cost $0.03 per million input tokens with free output. Nace retired Drex 1.5 and the first Drex Max on 8 November and answers them with Drex Max v1.1. OpenRouter serves Drex through DeepInfra. Nace reports 67.6 for Drex Max on the public Jev Decision Index 0.3.2, a vendor-run figure. On the independent Full score in edition 0.4.1, Drex Max scores 66.6, 41st of 301 in the 12 November file, against Jev 1.14’s 60.24, Drex 2.0 52.76 and Drex 1.5 45.78; Nace submitted the public runs. Benchmark Heaven’s JevBench measured Drex Max on its hosted API board at a Capability Score of 84.2.

Microsoft-Decision-1 from Microsoft

Microsoft-Decision-1 is Microsoft’s decision model, announced on 9 October 2026 and sold through Microsoft Foundry. Microsoft post-trained Qwen3.5-9B so that, given a fixed set of options, it returns a calibrated probability for each one. It answers yes/no, multiple-choice and rating questions, grades answers against a rubric, reads text only and takes up to 32,768 input tokens. Microsoft Learn publishes the request: a state and named questions in Jev’s shape, sent to <your-foundry-resource-endpoint>/providers/microsoft/v1/systemone with an Entra ID token or an api-key header. The weights are closed.

It costs $0.042 per million input tokens with free output, the same as Jev. Microsoft’s own 36-benchmark comparison gives it 83.5% average accuracy against 81.9% for Quyet-1.0-Large. It reports a median of 85 ms measured through Foundry, against 240 ms for Jev 1.13.0 taken from JevBench. The accuracy figures are vendor-run, the two latency figures come from different harnesses, and Jev has no accuracy score in that comparison. OpenRouter also lists it, served by Azure at the same price. It is on Benchmark Heaven’s JevBench and the Artificial Analysis Decision Index, but not the Jev Decision Index.

OpenAI Decisions API

The OpenAI Decisions API runs on GPT-6 Luna and was announced on 29 September 2026. You send text or images with typed questions, and it returns a probability for a predicate, a probability per option for a choice, or a weighted level for a score. It opened in public beta on 6 October 2026 and has been generally available since 22 October at POST https://api.openai.com/v1/decisions, for $0.10 per million input tokens with free output, and on OpenRouter at the same price. Its request format differs from Jev’s /v1/systemone. Every, which had preview access, found it ahead of Jev on some early tests and behind on its broader evals.

Span-02 from Respan

Span-02 is Respan’s general decision model, released on 15 October 2026 and a different model from Span-01. It answers Choice, Score and Noul questions, plus number questions that return a value, about text, JSON and up to four images. When it is unsure it can think briefly before answering, within a deadline you set per request. It costs $0.01 per million input tokens with free output, on Respan’s POST /api/v1/decisions endpoint, which takes the System One request format, and on OpenRouter. The weights, a fine-tune of Gemma 4 26B-A4B, are on Hugging Face with Respan’s own serving engine.

On the Jev Decision Index 0.4.1 it scores 59.07 on the Full score, 78th of 301, against Jev 1.14’s 60.24. Its calibration error of 0.114 is higher than Jev’s 0.0799.

mammoth-1 and elephant-1 from Baseten

mammoth-1 (27B) and elephant-1 (9B) come from Base Labs, the research team at Baseten, and were released on 21 October 2026. Both are post-trained from Qwen3.5 models, answer Choice, Score and Noul questions about text and images, and run on Baseten’s POST https://inference.baseten.co/v1/systemone, the same request shape as Jev. They were free on Baseten and OpenRouter until 27 October 2026. mammoth-1 now costs $0.10 per million input tokens and elephant-1 $0.035, on Baseten and OpenRouter. The weights are MIT on Hugging Face.

On the Jev Decision Index 0.4.1, mammoth-1 scores 68.44 on the Full score, 22nd of 301 in the 12 November file, and elephant-1 60.26, level with Jev 1.14’s 60.24. On the Artificial Analysis Decision Index v2.1, Elephant 9B scores 66.4, 10th of 22 decision models and just ahead of Jev 1.14’s 66.2, and Mammoth 27B scores 71.4, 4th. The Index row for elephant-1 records that its training included Decision Index train splits, as the authors declared.

SubQ Snap from Subquadratic

SubQ Snap is the decision model from Subquadratic, a research company that builds long-context models on its own sparse attention method. The first release came out on 12 October 2026, SubQ Snap 1.1 on 22 October and SubQ Snap 1.2 on 11 November. It answers Noul, Choice and Score questions about text and up to four images at POST https://api.subq.ai/v1/systemone, with Jev’s state and questions fields. One request can hold 384K tokens. It costs $0.04 per million input tokens with free output. It launched in public preview, and SubQ’s docs no longer call it a preview.

On the Artificial Analysis Decision Index v2.1, checked on 14 November 2026, SubQ Snap 1.2 scores 72.6, second of 22 decision models and level at one decimal with the leader, Drex Max v1.1. SubQ Snap 1.1 scores 67.9, and Jev 1.14 66.2. Benchmark Heaven’s JevBench gives the first release a Capability Score of 81.5 on its hosted API board. Subquadratic’s own run of the Jev Decision Index 0.2.1 suite gives 1.1 65.7 against 60.6 for Jev 1.14, a vendor-run figure. The multimodalart board has no row for it.

An LLM with structured outputs

The oldest alternative is the one most teams already have. Ask GPT, Claude or an open LLM for JSON that matches a schema, and let the provider enforce the schema. You get any label set you like and no new vendor. You also get LLM latency and LLM prices. There is no probability to threshold on, only token log-probabilities where the provider exposes them.

SOTAAZ ran one LLM on BANKING77, a set of 77 banking intents. GPT-5.6 Terra got 83.8% at a median of 1,351ms per message. That was better than every open System One option in the same test, and about 50 times slower than Laya. Jev vs GPT and Jev vs Claude work through cost per decision. Libraries like Outlines and Instructor give you the same guarantee on models you host yourself.

Open-source System One models

These ship trained weights you download and run. Kev-4B is the one exception with a paid hosted API, on OpenRouter since 25 September 2026.

Laya

Laya, from Convai Innovations, is an encoder with a decision head: ModernBERT-large for English at 421M parameters, and mmBERT-base for other languages at 322M. It answers Choice, Score and Noul. Convai measures 32.8ms to 39.5ms for one question on a Tesla T4 and gives 193ms to 464ms on CPU. It is the most-starred project in the space by a long way.

Accuracy depends heavily on the task. Convai’s own README calls Laya “a fast base to specialise, not a zero-shot decision engine.” SOTAAZ found it strong with few labels: 86.6% on TREC’s six question types in 23ms, and 94.5% on AG News. It fell to 37.0% on BANKING77 because all 77 labels share one token budget. Giving it 20 candidates shortlisted by a small embedding model raised that to 59.1%. On the Berkeley exam test it scored 31.2%, against about 30.1% for random guessing. Its calibration error is 0.466 as shipped and 0.081 after fitting a temperature, so fit one on your own data.

Kev

Kev is Jared Palmer’s family of LoRA adapters on Qwen3.5 at 0.8B, 4B and 9B, plus a 27B on Qwen3.8. It serves TypeSafe’s /v1/systemone shape, so the official typesafe-sdk Python client works against it. It runs on CUDA, ROCm or Apple Silicon through MLX. The author reports Kev-4B at 41.5ms to 145ms of model time on an L40S for a new state, and Kev-27B at 46.5ms to 178ms on a B200.

Kev publishes a direct comparison with Jev and says it is not a controlled one. On datasets Kev was not trained on, Kev-9B scores 0.822 on the development set against Jev’s 0.857, and Kev-27B scores 0.848. Kev-27B needs an 80 GB GPU and starts from Qwen’s post-trained release, whose training data is unknown. On knowledge questions the gap is wider: 0.74 on MMLU for Kev-9B and 0.84 for Kev-27B, against Jev’s 0.90. Training covered at most 384 state tokens, so long documents fall outside what it has seen.

Bespoke Nimble

Bespoke Labs’ Nimble fine-tunes Qwen3.5-9B with LoRA and answers choice and true-or-false fields. The team published its 2,676 training examples and 324 held-out examples. On those 324, Nimble matched 90.12% of reference labels, Jev 1.13 matched 93.21%, and untuned Qwen3.5-9B matched 66.36%. The README flags the limits: the labels are synthetic and come from six source families. The 9B weights need about 18GB before quantisation. Median latency was 106ms on an H100 and 444ms on an M5 Pro Mac, against 246.7ms for Jev through its API in the same table.

CLM

CLM takes a different route. It trains separate encoders for states and actions and scores them against each other, on a frozen Qwen3-8B with a small trained head. Because action embeddings can be cached, it is built for agent loops that see the same options again and again. The authors claim parity with Jev and up to 9x the speed on computer-use, gaming and tool-calling tasks. Nobody has reproduced that yet. It needs Linux and an NVIDIA GPU.

GLiNER2.5-Decide

GLiNER2.5-Decide is Fastino Labs’ entry, released 24 September 2026: a 340M-parameter encoder (Hugging Face’s own weight metadata reports 486M, an unresolved discrepancy) fine-tuned from Fastino’s gliner2-large-v1, a DeBERTa-v3-large base. It evaluates typed questions and rules against a passage in one pass and, beyond a typed answer, can also return spans, relations and structured records in the same call. Fastino’s own benchmark, an internal 17-dataset suite called “Fast Decisions,” gives it 60.1% average accuracy against 57.5% for what Fastino calls JevK5, 56.4% for SemIf and 46.6% for Laya, leading on 9 of 17 datasets. Latency, also from Fastino, is a p50 of 167ms on a 48-vCPU CPU and 38 to 47ms on GPU for short requests. None of these figures are independently reproduced. It does not serve TypeSafe’s /v1/systemone wire format; it loads through its own gliner2 Python package. Jev vs GLiNER covers it against Jev and against the original span-extracting GLiNER models in the same family.

Jeffan-31b-1710

Jeffan-31b-1710 is a full fine-tune of Gemma-4-31B-it, published on Hugging Face on 19 October 2026 by one developer, Trong Vu. It reads one multiple-choice question at a time in its own prompt layout and returns a probability for each option, so it does not serve the /v1/systemone request. On the Jev Decision Index 0.4.1 it scores 69.75 on the Full score, 17th of 301 in the 12 November file; it was second of 229 on 22 October. Its training data is undisclosed. The weights moved from CC BY-NC 4.0 to Apache 2.0 on 22 October 2026.

H2O-Lightning-31B

H2O-Lightning-31B is H2O.ai’s larger decision model, a Gemma-4-31B-it fine-tune published on 20 October 2026. It reads images and serves POST /v1/systemone through H2O’s shim on vLLM. Its v1.3.0 scores 69.92 on the Jev Decision Index 0.4.1 Full score, 16th of 301, and JevBench gives v1.0.7 an Intelligence score of 79.3. Its licence allows non-commercial, non-production use only, so unlike the Apache 2.0 H2O-Lightning-4B it cannot go into a product.

Von and Decider

Von is a 395M ModernBERT decision model that serves the same API and installs as von-sdk from pip or npm. Its own table puts Von 1.1 well behind Jev, at 72.0% on a 49-task suite where Jev scores 96.6%. Decider fine-tunes Qwen3.5-2B-Base, serves TypeSafe’s wire format and claims calibrated probabilities, but publishes no measurement.

Community clones and experiments

Most of the “jev open source alternative” results are here. These projects take an existing open model, frozen, and read the answer off its output scores. They train nothing new, or only a little. They show that the interface is easy to copy, and they show little about quality.

  • SemIf, formerly openjev, by TheoLeeCJ. It reads option probabilities from a frozen Qwen3.5-4B on one RTX 3090, with MLX and CPU paths too. On 21 yes-or-no questions about one state, it answered in 1.02s against 5.33s for the same model writing a JSON array. It scored 61.6% on the Berkeley exam test, well above Laya and well below Jev’s 83.7%.
  • OpenJev by razorback16 (repo). It turns DiffusionGemma 26B-A4B, a model that fills in all its output tokens at once, into a Jev-compatible server. It was the best open option on BANKING77 in the SOTAAZ test at 66.9%. Codiv hosts it free with 100M input tokens per account, billed as a public experiment, not a paid service. The Gemma team’s djev-run deploys a related DiffusionGemma server to Google Cloud Run, and open-jev by JoshuaSP is an earlier experiment on the same idea.
  • NanoJev, a 0.6B replica with weights, dataset and a full training pipeline. It is a good codebase to learn from. Its released checkpoint was trained on game decisions like Snake and mazes. It scored 18.4% on TREC and 20.1% on BANKING77 in the SOTAAZ test, so do not use it for text classification.
  • jevlike, a tiny byte-level model trained from scratch to pick among changing options, with Doom and chess demos. It is a research toy, not a classifier you deploy.
  • Rizzo Flow (repo) runs frozen Spark-X2.5 models through llama.cpp. That means Metal, CUDA, Vulkan or CPU from one install. It reports about 49ms per decision on an RTX 5060 Ti and says it is tied with SemIf on quality. It claims no lead over either project.
  • Laptop and Mac experiments: jevmlx, mini-jev and jev-on-a-laptop. Each one reads options off a small model on Apple Silicon. openjev-sglang is the opposite end: a 35B model on a B200.

The full list, with filters, is in the Jev alternatives directory.

Two corrections to older roundups. Latent.Space’s 19 September list describes Kev as a 0.5B adapter on Qwen2.5. The current family is 0.8B to 27B, on Qwen3.5 and Qwen3.8. The same list says Laya was trained with PPO. Laya’s model card says REINFORCE with a group-mean baseline.

Classifiers that were never System One models

If your label set is stable, these often beat everything above on cost and accuracy.

In the SOTAAZ test, a logistic regression on MiniLM embeddings scored 90.3% on BANKING77 and 90.4% on TREC, at under 7ms on a CPU. No open System One option came within 20 points on BANKING77. Jev itself scored 76.3% on BANKING77 in a separate test by nibzard that SOTAAZ cites. We have not checked that run ourselves. The catch is that the classifier needed labelled training data, and a new label means retraining.

  • Fine-tuned encoders: ModernBERT, DeBERTa and BERT. Jev vs BERT covers the tradeoff in detail.
  • SetFit: a usable classifier from a few dozen labelled examples.
  • Zero-shot NLI models: labels at request time with no training, but one forward pass per label, so cost grows with the label count.
  • GLiNER: the original GLiNER pulls entity spans out of text, not decisions about it. Fastino’s separate GLiNER2.5-Decide model, covered above, does answer typed decisions. See Jev vs GLiNER.

How to choose

  • You need a hosted API and only pick one option per question. Try Tev1 on Together. Test it on your data first, since its only published number is from a reused dev set.
  • You need a hosted API with Score and Noul. Stay on Jev, or try Solar Decide, d1, Decider 1, Mercury Decide, Clef, Clef-omni, pplx-decider, Kai, Celeris Decision, Drex, Microsoft-Decision-1, the OpenAI Decisions API, Span-02, mammoth-1, elephant-1, SubQ Snap or Kev-4B on OpenRouter. An LLM with structured outputs also works if you accept the latency. Solar Decide is in beta. Kev-4B’s accuracy figures are its author’s own, the hosted d1’s are Liquid’s own reproduction, Decider 1’s come from a benchmark meraGPT built, Mercury Decide’s are Inception’s own, and Solar Decide’s come from one third-party test on 48 questions. Clef, Clef-omni, Drex and pplx-decider also have Jev Decision Index 0.3 scores. Kai’s, Celeris Decision’s and Microsoft-Decision-1’s figures are Hanzo’s, Celeris’s and Microsoft’s own, as are the accuracy claims Nace and Cloudflare make for Drex and Clef-omni.
  • Your code calls Jev and you want another hosted API. Solar Decide, d1, Decider 1, Celeris Decision, Drex, mammoth-1, elephant-1 and SubQ Snap use the /v1/systemone request shape, Span-02 takes the same body at Respan’s /api/v1/decisions URL, Clef and Clef-omni take the same fields through Workers AI, and Kev-4B takes the same request as Jev. Span-01, pplx-decider, Kai and the OpenAI Decisions API do not. Mercury Decide uses its own /v1/decisions endpoint on Inception’s API. Microsoft-Decision-1 takes the same fields at a Foundry URL, /providers/microsoft/v1/systemone.
  • You need to classify images as well as text. Clef takes up to four images per request, pplx-decider takes images in the state array, Celeris Decision takes up to 8 base64 images, and SubQ Snap takes up to 4. Every image figure these vendors publish is their own, so test them on your own images.
  • You need to classify audio or video. Clef-omni takes up to four audio clips and two videos per request. Its numbers are Cloudflare’s own.
  • You need a long input in one request. pplx-decider accepts up to 262,144 input tokens. Perplexity’s vendor-run tests on 12 November 2026 put v1.3 at 7 seconds for about 135,000 tokens and 16 seconds for about 243,000. Drex Max v1.1 takes 262,144 tokens for the whole request, and Drex 2.0 a state of up to 131,072. Nace’s vendor-run median of 2.0 seconds from 32K to 128K tokens was for Drex 1.5, now retired. SubQ Snap takes 384K tokens per request, and Subquadratic says it answers 100,000 tokens in about a second, a vendor figure.
  • You want to score behaviors in agent traces. Respan aimed Span-01 at that job. It returns present, absent or not observable for each behavior, and its accuracy figures are Respan’s own. Respan says Span-02 has replaced it.
  • Your code already calls Jev and must move to your own hardware. Kev, Decider, Von, CLM and OpenJev all serve the same endpoint, so you change the base URL. Rerun your own evals, because every public test shows a quality drop.
  • You want the closest quality to Jev on your own GPU. Kev-27B, Nimble and Kev-9B publish the smallest gaps to Jev, all on their own test sets. On Jev Decision Index 0.3.1, pplx-decider-v1.1-27b scores above Jev and Kev 27B 1.33 points below it. Both are 27B models, and Kev-27B needs an 80 GB GPU. Check them on yours.
  • You need CPU only or very low latency with a few labels. Laya, or Von. Shortlist candidates first if you have dozens of labels.
  • You have labelled data and fixed labels. Train an encoder or a logistic regression on embeddings. It is cheaper, faster and, on the public evidence, more accurate.
  • You want to learn how these models work. NanoJev, jevlike and SemIf publish their full pipelines and results.

Measure calibration as well as accuracy. Confidence gating, acting only when the probability clears a threshold, is the main reason to use a System One model at all. Most open projects either disclaim calibration or fit a temperature afterwards, so a threshold tuned on Jev will not transfer.

FAQ

Is there an open-source version of Jev?

No. TypeSafe has not released Jev’s weights, architecture or training method. The open projects copy its API shape on other models. Laya, Kev, Nimble, CLM, Von and GLiNER2.5-Decide ship trained weights under Apache 2.0. SemIf, OpenJev and most other clones read answers off frozen open models. None reproduces RLCD, TypeSafe’s unpublished training method for calibrated probabilities.

Can I run Jev locally?

No. Jev is a hosted API only, with no weight download or self-hosted build. To run a System One model locally you run a different one. Laya and Von run on CPU. Kev, Nimble and SemIf run on Apple Silicon or one consumer GPU. CLM and OpenJev need a larger NVIDIA card, or you can try OpenJev free on Codiv.

What is the best Jev alternative?

It depends. For a hosted API, try Solar Decide, d1, Decider 1, Mercury Decide, Clef, Clef-omni, pplx-decider, Kai, Celeris Decision, Drex, SubQ Snap, Microsoft-Decision-1, the OpenAI Decisions API or Kev-4B, or Tev1 for Choice only. On a large GPU, pplx-decider-v1.3 and Kev-27B score nearest Jev or above it. For a few labels on CPU, Laya. For fixed labels with training data, a fine-tuned encoder beat every open option on BANKING77.

Are the open alternatives as accurate as Jev?

Some come close, most do not. The broadest independent test, Jev Decision Index 0.3.1, gives Jev 60.11 on its Full score. One open-weight model scores higher, pplx-decider-v1.1-27b at 62.75, and Kev 27B is 1.33 points behind. Kev 9B v2 scores 43.27 and GLiNER2.5-Decide 9.24. On the Berkeley exam set, Jev scored 83.7% and SemIf 61.6%. Fastino’s own benchmark puts GLiNER2.5-Decide ahead of JevK5.

Which alternatives work with the TypeSafe SDK?

Kev, Decider, Von, CLM, OpenJev, openjev-sglang and Rizzo Flow serve TypeSafe’s POST /v1/systemone request and response, and so do Solar Decide, Celeris Decision, Drex, mammoth-1, elephant-1 and meraGPT’s Decider 1, unrelated to the open Decider. Liquid’s docs use the SDKs with d1 at https://api.liquid.ai. Change the base URL and code runs unchanged. Tev1 and Span-01 do not. Tev1 uses Together’s own API and returns one letter, and Span-01 uses /api/v1/scores.

Related guides