System One Models & Typed Decision Engines

by TypeSafe AI

The hosted model that introduced the System One name and the choice, score and noul question types. Generally available and billed per token.

by Liquid AI

Liquid AI launched the hosted d1 on September 29, 2026 with a free tier (model name d1:free). On October 5 it released open weights on Hugging Face: d1-3B (text and images, built on LFM2.5-VL-3B, 32K context) and d1-omni-600M (text with images or audio). Liquid reports d1-3B at 48.57 on Decision Index 0.2.1, the best score under 10B (self-reported).

by OpenAI

Announced at DevDay on September 29, 2026 and in public beta since October 6: $0.10 per 1M input tokens, with no charge for output or cache. Returns probabilities and a confidence for choice and score answers; OpenAI says it answers about 10x faster than the Responses API.

by Microsoft

Announced October 9, 2026 and available in Microsoft Foundry (Preview in the catalog) at $0.042 per 1M input tokens, the same as Jev, with output free. Microsoft reports the highest accuracy across 36 benchmarks and about 35x lower median latency than GPT-6 Sol (self-reported); Clef is not in its comparison, and Jev was added only in a later update.

by Cloudflare

Released October 1, 2026 with open weights on Hugging Face and hosted on Workers AI. On October 9 Cloudflare added Clef-Omni (audio input, $0.15 per 1M input tokens) and cut Clef-Flash to $0.038, below Jev, with a 24,576-token hosted context. Cloudflare reports Clef leading 7 of 10 decision benchmarks against Jev, and Clef-Flash at 38.8 ms median latency (self-reported).

by Convai Innovations

Non-autoregressive choice, score and noul heads on ModernBERT and mmBERT encoders, with a Router, fine-tuning and many community runtimes.

by Jared Palmer

Small Jev-like models on Qwen3.5 and Qwen3.8 bases. Each checkpoint ships a fitted temperature, and a fine-tune and deploy loop runs on Modal.

by Manjunath Shiva

Distilled from two open teacher models. The 400M nano model scores 0.796 on the typed-decisions benchmark, against 0.766 for Laya's fine-tuned checkpoint and 0.754 for Jev (author's measurements). Early release, updated often.

by Mapika

One-pass typed decisions with calibrated probabilities, trained on public data and labels from a local Qwen3.5-27B teacher. Choice supports 2 to 255 options.

by wfzyx

Built on ModernBERT-Large. Version 1.2 scores each option independently, so the answer no longer depends on the order the options are listed in.

by Bespoke Labs

Publishes the data curation, training and serving recipe behind Bespoke-Nimble-9B, which now takes 8,192-token inputs and up to 255 choices per field.

by TheoLeeCJ

Scores typed options directly from a frozen open model with no answer text or JSON parsing, with per-workload temperature calibration.

by Rizzo AI Academy

A local take on Jev with a LoRA fine-tune of Spark-X2.5. The project reports 0.648 on typed-decisions against 0.727 for Jev, and notes that probabilities are uncalibrated unless you calibrate them on your own data.

by Nokia Applied Research

Turns an existing LLM into a Jev-style decision model by reading its hidden states, with debiasing and calibration levels and no task-specific training.

by TianyuCodings

A compact replica with an end-to-end training pipeline, evaluated on game tasks such as ViZDoom, a maze and Snake against Jev.

by Together AI

An experimental supervised fine-tune of Qwen3.5-4B released on September 23, 2026, with the data recipe and a guide to training a similar classifier.

by Fastino

Takes label sets at call time and scores several heads in one pass. Built for operational decisions such as intent, routing, sentiment and priority.

by iApp Technology

A Qwen3.5-0.8B text tower whose output head is replaced by a 256-way slot softmax, continued-pretrained on Thai.

by mmastrac

Reads Jev-style decisions from DiffusionGemma, a discrete diffusion model, in one denoise step through vLLM. The djev-dev fork adds native image inputs and live camera sampling.

by akhilaaa3

A decision classifier for text, images, audio and video, fine-tuned on 30,000 questions: a probability per option instead of an explanation. Reports 86.15% on matched JevBench groups (self-reported).

by Zefan Cai

Scores the candidates you supply for choice, noul and score questions without generating an answer, using a trained head on top of Qwen3.5.

by alibiserikbay

Qwen3.5-4B with a merged LoRA, read out with the SemIf protocol and one calibration temperature. Questions with more than 16 options are read in several passes.

by kotoba-lang

An independent reproduction of the Jev interface: choice (up to 255 options), score (2 to 10 levels) and noul in one forward pass, measured on public gold labels.

by malevrigns

The language-model head is removed and a candidate head scores the options. The published checkpoint judges whether a coding task is actually finished.

by chaoliangUNSW

GGUF decision models that return calibrated option probabilities from a single token. v3 reports 79.2% on 2,000 typed decisions and takes 25,600-token inputs.

by ToufiqQureshi

A 68M encoder that beats Jev on the public typed-decisions benchmark (74.0% vs 72.7%) and comes within 2.6 points of Laya at a sixth of the size, with the lowest calibration error of the three (author-reported).

by PostHog

Thinks before it answers: 0.935 on the public JevBench items vs 0.866 for Jev (0.865 vs 0.730 on the hard tier), at about 0.3 s per request without thinking and a 3.3 s median with it on one H100. Jev still leads on MMLU-Pro (author-reported).

by ejhshen (Junhao Shen)

Options and criteria are inputs, so the output follows whatever candidates a request supplies, and reordering them does not change the answer. Trained with SFT then RL on the public OpenJevData-140k set; training code included.