All 130 System One / decision models we track, with maturity and capability scores, latency, per-token price and license — hosted Jev plus open peers.
Grounded AI
LoRA adapter
Grounded AI’s tool-call verifiers for agents: a 9B LoRA on Clef-flash (Apache-2.0) and a 3B on Liquid d1-3B (LFM Open License). Author-run results only.
Command Code
Open Gemma-4-31B decision fine-tune from Command Code with a local /v1/systemone server. Self-reports 58.15 on Decision Index 0.2.1. Apache-2.0.
Ainvest
Ainvest's Apache-2.0 text decision model, a weight merge on Qwen3.8-27B with a 255-way readout. Decision Index 0.4.1 Full 68.23 (#10), ECE 0.042.
andyzhang232 (independent)
LoRA adapter
Independent Jev-style LoRA for Gemma 4 31B: yes/no, choice and score with a probability per option. Decision Index 0.4 Full 69.05 (#5).
Zaid Mukaddam (Scira)
Four Apache-2.0 decision models, 0.5B to 27B, that read photos and return boxes. Free API. Board (17 Oct): Decision Index 0.3.2 Full 64.08 (#6).
Arc53 (DocsGPT)
Small MIT decision models from Arc53 for RAG and agent apps (relevance, injection, grounding, routing), 0.6B and a CPU-first 0.4B. Used in DocsGPT. Author numbers only.
denis-pplx · hosted by Perplexity
Self-host via /v1/systemone, or hosted as Perplexity's pplx-decider-v1-27b (same weights) at $0.02/M input. Board (17 Oct): JevBench Intelligence 55.6.
Remek Kinas (basal)
basal: a Polish/English typed-decision family (basal-2.0 on Bielik, basal-omni on Gemma 4) with an Apache-2.0 engine. basal-omni-1.0-ultra is #1 on JevBench open-weights Capability (82.5).
avbiswas (independent)
Tiny (≈0.4B) open System One decision model: 20 Qwen3-0.6B layers plus an attention decision head. Noul / Choice / Score, order invariant. Weights are non-commercial (CC-BY-NC-4.0).
Pixilab
A fast, calibrated decision model from Pixilab on Gemma-4-26B-A4B-it. Board (17 Oct): Decision Index 0.3.2 Full 57.76 (#32).
sanghwa-na (independent)
Bobcat 1.3: an independent Gemma-4-31B decision model (Apache-2.0, EN/KO, System One HTTP shapes). Full 62.83 on Decision Index 0.3.2 (#10, Jev 60.11); median 64 ms on the board's GPU.
cbuk10001 (independent)
LoRA adapter
Cam1: an independent Qwen3.5-4B LoRA plus a Clef-derived decision head (Apache-2.0, 20k training rows published). Decision Index 0.3 public 30.72, about half of Jev. Needs a custom FP32 runtime.
Celeris
Celeris's closed diffusion-model decision API: text, data and images, $0.04/M input. Fastest on the Artificial Analysis index (63 ms) but mid-field on accuracy: AA 52.3, JevBench API #17 (62.0).
Apertix AI
Apertix AI's open Jev-style decision family (27B, 9B, 4B, Apache-2.0) with calibrated probabilities in one pass. Decision Index 0.4 Full 62.64 (#26) for the 27B.
Ninh Nguyen (xoai, independent)
LoRA adapter
Independent Gemma 4 LoRA family with a /v1/systemone server: choice, yes/no and score. Chon-1.0-12B leads JevBench open weights (Capability 85.3).
Cloudflare
Cloudflare's open multimodal decision models (Clef 27B, Clef-flash 9B, Clef-omni 30B-A3B, Apache-2.0) on Workers AI at $0.24 / $0.038 / $0.15 per M input, and in Ollama.
FluidInference
Open 0.70B typed decision model distilled from Clef-flash into Qwen3-0.6B, with PyTorch weights and Core ML for Apple devices.
Contrastive-LM (research)
Open contrastive state/action model: Qwen3-8B + projection heads, typed Choice/Noul/Score API, relative probabilities. Latency p50/p99 not published; 0 is an unavailable-data sentinel.
Cribl, Inc.
LoRA adapter
Cribl's open decision adapters on Qwen3.5 (0.8B to 27B, Apache-2.0): choice, noul and score from one forward pass. Built for telemetry routing and triage. Author numbers only so far.
tardellirs
A 149M typed-decision encoder on the official Decision Index 0.3 board (22.21, best at ≤150M). 20 ms on an A100. Weights CC BY-NC 4.0: non-commercial only.
Liquid AI
Open 3.1B decision model from Liquid AI, fine-tuned from LFM2.5-VL-3B. Text and images, noul/choice/score in one pass. LFM license, not the hosted d1 API.
Liquid AI
Experimental 587M decision model from Liquid AI. Text plus image, or text plus up to 30s of audio. Same question types as d1-3B, much lower self-reported index.
Liquid AI
Liquid AI's larger hosted decision model: text and images, Noul/Choice/Score with zero output tokens. #1 on the Artificial Analysis Decision Index v1.0 (65). Weights closed.
Mapika (independent)
Open System One-style decision model: full FT of Qwen3.5-2B + RL calib v10. TypeSafe API-compatible; noul maps to boolean.
Mapika (independent)
First open model to beat Jev on JevBench — 4B dense with no RL stage, faster and cheaper.
arczhi (independent)
Open 1.7B decision model (Qwen3 fine-tune) with explicit candidate-scorer architecture: calibrated probabilities over arbitrary runtime candidates in one forward pass.
vLLM Semantic Router (vllm-sr)
vLLM Semantic Router's second open decision family: six Apache-2.0 sizes from 0.6B to 27B, Choice / Noul / Score in one pass. Board (17 Oct): Decision Index 0.3.2 Full 55.88 (#42).
FlyMy.AI (@denti)
MiniCPM5-2B based decision model with pointer head — evaluation preview, not production release.
vLLM Semantic Router
vLLM Semantic Router’s third open decision family: six Apache-2.0 sizes (0.6B–27B) that read text, images and video. d3 27B claims 65.45 on the DI 0.3 public suite.
FlyMy.AI (@denti)
LoRA adapter
Board (17 Oct): JevBench Intelligence 6.1.
vLLM Semantic Router (vllm-sr)
Small open encoder decision model (572M, Apache-2.0) from the vLLM Semantic Router project. Noul / Choice / Score in System One format, very fast, weak on the Decision Index.
vLLM Semantic Router (vllm-sr)
Open 9B decision model (Qwen3.5-9B, Apache-2.0) from the vLLM Semantic Router project.
Lukitaduarte (independent)
150M Apache-2.0 encoder for typed decisions in English and Portuguese on a CPU, with tool and calculator experts. Decision Index 0.4 Full 16.08 (#172).
mmastrac (independent)
DiffusionGemma 26B-A4B NVFP4 + structured decision server (DGX Spark/GB10 or Cloud Run via djev-run); System One API. Hardware-oriented stack.
Nace.AI
Nace.AI's open decision models: Drex Max v1 (27B, multimodal, weights since 30 Oct) and Drex v2.0 (35B-A3B). DI 0.4.1 Full 66.60 for Max. Free under $1M revenue only.
Upright (Upright Project)
Upright's hosted consequential decision model: per-stakeholder expected value plus a SystemOne-compatible API for choice, score and noul. $2.50/M input. Not on any public board yet.
Caio Vicentino (independent)
Open MIT decision models from Brazil (4B and 27B Qwen fine-tunes): option-letter readout, noul/choice/score, /v1/systemone server. Finance and trading focus.
Manu Romero (mrm8488)
A 0.3B Apache-2.0 fine-tune of EmbeddingGemma 2 that scores choice, score and boolean questions from a schema. The author reports 76.4% on typed-decisions.
SberAI (УЭСМО) / ai-forever
SberAI's open 823M encoder decision model for Russian (and English): Choice / Score / Noul / ranking in one encoder pass, MIT, with an int8 CPU build and a vLLM plugin.
plantwaterer (independent)
LoRA adapter
Independent LoRA plus Bayesian head for Gemma 4 31B that answers choice and yes/no questions with a probability per option. Decision Index 0.4 Full 61.69 (#31).
vixhal-baraiya (independent)
Independent open System One fine-tune of Qwen3-4B: Choice up to 256, Score and yes/no via a branching cross-encoder. Calibration RL claimed; no public numbers yet.
AutoTrust AI (not TypeSafe)
LoRA adapter
AutoTrust AI's open decision adapter on Gemma-4-26B-A4B, with adaptive thinking. Same weights as JEV-Gemma4-26B-A4B under a new name. Distilled lineage; not affiliated with TypeSafe.
Fastino Labs
Fastino's decision model, open weights since v1.2 (22 Oct 2026) on Qwen3.8-27B. Decision Index 0.4.1: #1 (71.52). API $0.04/M input.
Fastino Labs
Open Apache-2.0 encoder decision model from Fastino: schema-driven, multi-head classification, CPU-first, and no generated tokens. Latency is not published; 0 is an unavailable-data sentinel.
kouhxp (independent)
A 0.8B open decision model for CPUs: full fine-tune of Qwen3.5-0.8B shipped as GGUF, yes/no / choice / score with a reject option over a /v1/systemone runtime. Apache-2.0.
H2O.ai
H2O.ai's open 4B decision model (Qwen3.5-4B, Apache-2.0, text and images): #1 open-weights composite on JevBench (72.5, Jev 1.13 71.5), 29 ms measured p50, runs on stock vLLM.
HopitAI
Open LoRA adapter on Qwen3.5-4B for JevBench-style decisions. One forward pass, no generation. Author discloses benchmark-directed development.
mohit67890 (independent)
LoRA adapter
Open multimodal Jev-style LoRA on Qwen3.5-4B: photos + state in, typed choices with probabilities and a trained can't-tell out. Author cites #1 on JevBench and Image JevBench.
ZooWork (Serendipity One Inc.)
ZooWork's open 4B decision model: merged LoRA on Qwen3.5-4B, /v1/systemone runtime, hosted API. Self-reported JevBench 85.71% with the public set used for selection.
frontier-infra / getainode
Open decision model in 27B and 9B (Apache-2.0) with GGUF, Ollama and MLX builds: choice, noul and score in one pass. Decision Index 0.4 Full 65.37 (#14).
PostHog
LoRA adapter
PostHog's open 9B decision model that reasons before it decides: Qwen3.5-9B + LoRA + pointer head, trained with SFT and CISPO, Jev-compatible /v1/systemone server and SDK.
Gestalt Lab (independent)
LoRA adapter
Open LoRA on Qwen3-4B-Instruct-2507 for Jev-shaped choice/noul/score; author FEVER-family comps vs Jev are UnverifiedClaim.
tattrongvu (independent)
Independent Gemma 4 31B fine-tune that answers multiple-choice decision questions with a probability per option, one question at a time. Decision Index 0.4.1 Full 69.75 (#2).
TypeSafe AI
First public System One model — typed decisions with calibrated probabilities for software, not chat.
akhilaaa3 / Akhila Datta (independent)
Open multimodal System One–style classifier (Gemma 4 12B) — text/image/audio/video → typed options + probs; independent Apache-2.0.
chaoliangUNSW (independent)
Open Apache-2.0 0.8B decision model (full fine-tune of Qwen3.5-0.8B): noul/choice/score with fitted temperatures, 25.6k-token inputs, local /v1/systemone server, MCP and agent skills.
National University of Singapore (Qi Li, Xinchao Wang)
NUS decision-native model on a Qwen3.5-4B backbone: Choice with an explicit none, Noul and Score. Apache-2.0 weights since 18 Oct. Paper results are the authors' own; not on a public board yet.
HIT-TMG (HITsz-TMG)
Open embedding-based decision family from HIT-TMG: LoRA-tuned embedders for Choice/Score/Noul, with a Python CLI and optional Jev-shaped HTTP server. Latency unpublished; 0 is a sentinel.
alibiserikbay (Independent)
Open-weight Jev alternative on Qwen3.5-4B. Board (17 Oct): Decision Index 0.3.2 Full 41.54 (#91).
vinnylarouge (independent)
Research starter — train a small one-pass option scorer (Jev-like I/O shape); not TypeSafe and not a full System One suite.
Jurin AI
Japanese business decision models from Jurin AI (Tokyo), built on StartLux-Decision. 0.8B and 2B weights are CC BY-NC 4.0; 4B and 9B stay private. Vendor numbers only.
Supersonic Labs (Brazil)
Open 144M Apache-2.0 decision model from Brazil, fine-tuned from mmBERT-small: state + question + 2–20 options → one pick with probabilities. Runs on CPU.
Sqwish Labs
Sqwish Labs' decision-model family: fat and chub read images, core is text-only, on the OneHot API. Fat is #1 on the JevBench API board (77.8) and ImageJevBench (75.9). Core/dot weights unverified.
Jared Palmer (independent)
LoRA adapter
Open family of Jev-compatible decision models (Kev 1.0: 0.8B, 4B, 9B, 27B; Apache-2.0). Pointer head on Qwen backbones; GPU and Apple Silicon (MLX) server; on OpenRouter.
PelaAI
KnowLine-35B-A3B Gen1: PelaAI's Apache-2.0 Qwen3.5-35B-A3B decision model (EN/ZH). Full 62.92 on Decision Index 0.3.2 (#9, Jev 60.11) but public 69.49 (#2); ECE and latency are self-reported.
ConvAI Innovations
Open ModernBERT-large System One–style decision model (~421M) with typed answers and probabilities.
Interfaze (YC P26)
LoRA adapter
Open LoRA on Qwen3.5-4B — TypeSafe-compatible System One API, 68.9% S1Bench macro, calibrated probabilities, zero output tokens.
OpenLAIR (Dingjie Song, Lichao Sun et al.)
LoRA adapter
Looped decision models (Apache-2.0): one checkpoint answers at depth 1–4 with a probability per option. Lev-2.6B: 63.2% → 76.2% on JevBench public (authors). Not Interfaze's Lev.
PNPM244 (independent)
31B open decision model on Gemma 4 31B, English and Vietnamese. #2 on Hugging Face's Decision Index 0.3.2 Full score (65.61), with the private tasks run by the maintainers.
Liquid AI
Liquid AI's first decision model: hosted API, Noul/Choice/Score with probabilities, TypeSafe-SDK compatible. Claims to beat Jev on HF's Decision Index (self-run). Weights closed.
FrontiersMind (independent)
Open Apache-2.0 decision-model family, 0.1B to 27B (the 27B is a full fine-tune of Qwen3.8-27B): typed noul/choice questions in one pass; pip package + /v1/systemone server. Author-reported numbers.
LycheeAI
Lychee-27B: a Qwen3.8-27B decision model under CC BY-NC 4.0 (non-commercial) with a /v1/systemone server. Full 62.99 on Decision Index 0.3.2 (#8, Jev 60.11); no calibration or latency data.
Base Labs (Baseten)
Base Labs' open decision models on Qwen3.5 (27B and 9B, MIT): choice, yes/no and score in one pass. Hosted on Baseten and Vercel AI Gateway. DI 0.4.1: 68.44 and 60.26.
Maincode
Maincode's open ~26B decision models (Apache-2.0, AU). Board (17 Oct): Decision Index 0.3.2 Full 59.23 (#22).
thelion.ai (lion-ai)
LoRA adapter
Open medical decision-model family (4B–31B adapters) from thelion.ai. Strong author robustness numbers vs Jev 1.13; CC BY-NC 4.0, not for commercial or clinical use as is.
sky7350 / akivet (independent)
Small open decision model: merged LoRA on Qwen3.5-4B, English + Korean, Noul/Choice/Score via /v1/systemone, GGUF builds. JevBench public outputs published.
Microsoft
Microsoft’s closed Qwen3.5-9B decision model in Foundry. JevBench v1.6.1 ranks it sixth among hosted APIs, with strengths in several use cases and a lower overall score than Jev.
r-ms (independent)
Teaching experiment: letter-logit reads on frozen Qwen3-4B for closed choices — not a trained decision model. score not measured.
MISA JSC
MISA JSC's open Gemma-4-31B decision model (Apache-2.0, English and Vietnamese, text and images), served on /v1/systemone by SGLang. Published 16 Oct with no benchmark numbers.
Nimruz
Persian-first open decision models from Nimruz (Apache-2.0, GGUF): v3 3B for text, v2 4B with images. /v1/systemone body. Author-run results only.
MoganAI
LoRA adapter
MoganAI's non-commercial LoRA for Gemma 4 31B: choice and yes/no with a probability per option, no generated text. Decision Index 0.4 Full 61.06 (#35).
MoLeMo-Lab (academic)
Open 0.85B multimodal System One model (Qwen3.5-0.8B base) — wire-compatible with TypeSafe SDK, 262k native context, MIT license, ECE 0.79%.
TianyuCodings (independent)
Open 0.6B parallel decision model (Qwen3-0.6B + decision heads) — independent Jev-style replica, not TypeSafe.
TokenRhythm
TokenRhythm's 4B decision model: Kev-style pointer head trained on NeoHorse-1-4B, Noul/Choice/Score, vLLM/SGLang adapters. Official Decision Index: 36.75 (#30), vs a self-reported 77.70.
Bespoke Labs
LoRA adapter
Open LoRA on Qwen3.5-9B — typed decisions + probs; open Jev recipe, not a frontier base model.
Bespoke Labs
Bespoke Labs’ closed hosted decision model: nimble-v4-max at $0.04/M input (v4 now aliases it), output free. JevBench API: composite 66.2 (#14), capability 79.7; vendor DI claims unverified.
plantwaterer (independent)
Independent Apache-2.0 fine-tune of Gemma 4 31B with a calibrated head for choice and yes/no questions. Decision Index 0.4.1 Full 68.10 (#11), ECE 0.040.
OmniJev (open source)
Open multimodal System One decision family (0.8B-27B, Apache-2.0): Qwen3.5 / Qwen3.8 fine-tunes for Noul / Choice / Score on text, images and video. 1.5-27B: Decision Index 0.3.2 Full 59.90.
kotoba-lang / com-kotobalabs (independent)
Open DeBERTa-v3-large typed-decision encoder (choice/score/noul) on public gold; ~28ms H100 / ~1.8s M1 CPU (author).
manjunathshiva (independent)
Open calibrated System One family (nano ~400M, small 4B, medium-td 30B-A3B): choice/score/noul via pip install opendecider. Author head-to-head vs Jev/Laya = UnverifiedClaim.
Alex Wortega (independent)
Open Qwen3.5 NLI cross-encoder framed as a Jev-style decision model; 0.8B long-context demo + multimodal I/V/T v2, RLCD note, SGLang path; openjev.co coming soon.
OpenJev project (independent)
Open 27B decision model (Qwen3.8-27B fine-tune) for text, web pages and screenshots, with a /v1/systemone helper and llama.cpp support. Weights CC BY-NC 4.0.
iApp Technology
Open 0.8B Thai + English System One model from iApp, with an official Ollama build for /v1/systemone. Apache-2.0, about 10.5k downloads.
Phocinae
A 144M English/Chinese encoder decision model on mmBERT-small (Apache-2.0) for local routing and escalation. Strong on its own typed-decisions run, weak on JevBench public (54.5%).
masterset-ai
Image decision model (Darwin-35B-A3B base, Apache-2.0) for visual inspection and QC: choice, yes/no and score over up to 4 images. Small author-run tests only.
Perplexity
Perplexity's second open decision model: Qwen3.8-27B with noncausal attention and a 255-row readout. #1 on Decision Index 0.3; hosted at $0.02/M input.
Perplexity
Perplexity's v1.2 open decision models: 27B (Qwen3.8) and Flash 4B (Qwen3.5), joint requests, Apache-2.0. Full 65.55 on Decision Index 0.3.2 (#4); API $0.02/M and $0.01/M input.
Chinh Nguyen (chinhnc)
Open decision-model family from Chinh Nguyen: Gemma-4-31B, Qwen3.5-4B and three encoder sizes. Large is #1 on JevBench v1.6.1's open-weights board (71.4).
kshetrajna12 (independent)
LoRA adapter
Open LoRA decision adapter on Qwen3.5-4B with System One-shaped API; ~100ms cached author path; WebGPU 0.8B demo.
Shanghua Gao (independent)
Open Apache-2.0 2B decision model (Qwen3.5-2B-Base + trained scoring head): first RSI-Jev release with a listwise ranking RL stage that stuck. Local /v1/systemone server.
Levanto Labs
Levanto Labs' closed decision API: yes/no, choice, scale, sort and tags with calibrated probabilities. #1 on Benchmark Heaven's JevBench v1.6.1 API board (composite 74.0, capability 78.6).
Minnesota NLP (minnesotanlp)
LoRA adapter
Looped decision adapters on Ouro (Apache-2.0): one question, 2–26 options, a probability after every loop. Authors report 75.8% for the 2.6B after eight loops.
Sthanika AI
LoRA adapter
Sthanika AI's open decision-adapter family (0.8B–31B) on Qwen3.5 and Gemma 4. Sieve 27B scores 61.25 on the official Decision Index 0.3 (rank 6).
Respan AI
Closed, waitlisted reasoning classifier from Respan: RLAIF-trained, one forward pass, parallel probabilities per written behavior definition — but no open weights and no public API contract.
Respan AI
Respan's open-weight decision model: Gemma 4 26B-A4B fine-tune with its own engine, Choice/Noul/Score/number, images, and a thinking deadline set per request.
sparrow-decider (unofficial community project)
Community family of tiny decision models, 47 MB down to ~2 KB, distilled from Strands Decider: Noul / Choice / Score on CPUs and ESP32-S3 microcontrollers. Apache-2.0, no board row.
StartLux Labs
StartLux-Decision v1 and v1.1 (0.8B-27B; Instant and AutoThink), CC BY-NC-4.0, with a /v1/systemone server. The v1.1 27B is #1 on the official Decision Index 0.3.2 (Full 69.12).
AWS (Strands Labs)
LoRA adapter
AWS Strands Labs' open 2B decision model: Qwen3.5-2B with a pointer head and a rank-16 LoRA, Noul / Choice / Score over /v1/systemone. Weights, data and scripts under Apache-2.0.
Subquadratic
Subquadratic’s closed long-context decision API: 384K tokens of text and images, $0.04/M input, output free. Snap 1.1 (22 Oct): AA Decision Index 65, tied #1.
Singularity Compute
Open-weight (Apache-2.0) DiffusionGemma 26B-A4B decision model from Singularity Compute; #1 on the JevBench API board (78.5) and hosted on ASI:Cloud at $0.03/M input.
Invergent (surogate)
Invergent's open decision model: Gemma 4 26B-A4B full fine-tune, text + images, Noul/Choice/Score. Apache-2.0. Board (17 Oct): Decision Index 0.3.2 Full 57.43 (#35).
Damien Sileo / tasksource
Open Apache-2.0 ~149M multi-vector decision model (LateOn fine-tune): choice/noul/score via ColBERT late interaction, not a classifier head. Latency unpublished; 0 is a sentinel.
Together AI
Open 4B Jev-inspired classifier from Together AI: Qwen3.5 SFT with typed choice prompts, but autoregressive letter output and no calibrated probability distribution.
FLock.io
MIT ~1.88B typed decision model (hybrid Qwen3.5/DeltaNet) adapted from decider-2b — one forward pass, zero output tokens, declared options + calibrated probs.
HAL-X AI
HAL-X AI’s 322M multilingual encoder decision model (mmBERT-base, RLCD, Apache-2.0), tuned for Azerbaijani support tickets. Author-run numbers only.
lostargon (independent)
Open 0.6B System One–style decision model (Qwen3-0.6B LoRA merged) — Choice / Score / Noul with calibrated probs; not TypeSafe.
Torchcast AI
Gemma-4-12B-it LoRA decision fine-tune from Torchcast AI, #2 on JevBench v1.6.1's open-weights board (69.9). Weights are CC BY-NC 4.0: non-commercial only.
MachineFi Labs
MachineFi's hosted decision API, now v1.3: Choice, Score and Noul on /v1/systemone plus image and sampled-video input. Closed weights, lineage undisclosed. Vendor-submitted, vendor-run numbers.
Turion AI
Gemma 4 31B decision model from Turion AI (Apache-2.0): choice, yes/no and score with a probability per option, /v1/systemone server included. Decision Index 0.4.1 Full 70.47 (#4).
Valen-Team / Liuziyu77 (independent)
Open multimodal System One model: text, images and video in; Choice/Noul/Score probabilities out. Qwen3.5-2B + trained decision head (SFT + experimental RLCD).
nandakishorm
Vega 0.8B / 4B: a frozen Qwen3.5 backbone with a small trained decision engine, 73K context and images. Licence is a Hub tag only (Apache-2.0). Independent JevBench: 4B Intelligence 16.7.
Heman10x-NGU (independent)
~151M ModernBERT-base open decision model (choice/score/noul); dual-channel calib claims; Jev/Laya comps UnverifiedClaim.
victorbennett (independent)
Independent 4B decision model with its own engine for CUDA, CPU and MLX and a /v1/systemone server; non-commercial weights. Decision Index 0.4 Full 49.31 (#71).
JudicialMind
Legal-first decision models from JudicialMind, 0.6B to 27B, on fine-tuned Qwen3 / Qwen3.5. Apache-2.0 for now. The author's tables beat Jev on contract tasks, not overall.
MCG-NJU (Nanjing University)
A System One model for video: replaces autoregressive answer decoding with a shared scoring head that returns temporal grounding spans and multiple-choice QA probabilities. Weights on Hugging Face.
Rover00 (independent)
Independent Apache-2.0 multimodal decision model on Qwen3.8-27B: choice, yes/no and score over text and images. Decision Index 0.4.1 Full 68.49 (#8), ECE 0.051.
Eldan Ring (independent)
Open Gemma-4-12B-it decision fine-tune with a local /v1/systemone server. #3 on JevBench v1.6.1's open-weights board at 68.9 (Q8). Apache-2.0 GGUF weights.
RBDash-Team
RBDash-Team's 27B multimodal decision model (Qwen3.8 VL, Apache-2.0), formerly RBDash-Decision-v0.2: #1 on the Decision Index Vision board (Full 73.3). No text-board row.
Juspay / Xyne
Open source multimodal Jev-like model by Juspay — Qwen3.6 based, 260k context window, text + visual classification; enterprise-grade with data sovereignty focus.
No models match these filters.
