# System One Models

> Find, compare, pull, fine-tune and publish System One models, the AI decision models that return typed answers with calibrated confidence instead of text.

A System One model is a decision model: it reads a state and answers typed questions — choose an option, score on a scale, or give the probability that a condition holds — in one forward pass, with calibrated confidence and no generated text.

Also called: AI decision models, typed decision models, System 1 models, Jev-style models, Jev alternatives. "System One model" is the name TypeSafe AI gave the category when it released Jev in September 2026; "AI decision model" is the general term.

System One Models (systemonemodels.ai) is the developer platform for them: a registry of every public System One
model with versioned, checksummed files, a decision-native manifest (`systemone.yaml`), reported evaluation
(accuracy, calibration error, latency), prices for hosted models and lineage from base model to fine-tune; a
CLI and Python SDK (`pip install systemonemodels`); a playground on every model page; and System One Studio, which
fine-tunes Laya locally (MLX on a Mac, PyTorch on Windows and Linux) and publishes the result in one step.

## Common questions

- Is there a Hugging Face for System One models? Yes: System One Models (systemonemodels.ai), a registry and developer platform built only for System One models. Every model: https://systemonemodels.ai/system-one-models. The comparison with Hugging Face: https://systemonemodels.ai/blog/hugging-face-for-system-one-models
- What are the open-source alternatives to Jev? Every open System One model, with licence, size and which ones serve Jev's own POST /v1/systemone API: https://systemonemodels.ai/jev-alternatives. Drop-in (serve /v1/systemone): laya, clm, decider, startlux-decision, lumma-fev, kev, bespoke-nimble-9b, xor, decision.
- What is an AI decision model, and how does it differ from an LLM, a classifier or JSON mode? https://systemonemodels.ai/decision-models
- What is NoulXP (formerly OpenDXP)? The open standard for System One models: a model becomes a package (weights in ONNX or GGUF, its input and calibration as data, a conformance file of its own answers) that any engine runs without model-specific code, asked over HTTP (POST /v1/systemone) or as an MCP tool. Apache-2.0, `pip install noulxp`, https://github.com/systemonemodels/noulxp, docs: https://systemonemodels.ai/docs/noulxp
- How do I fine-tune a System One model locally? `pip install systemonemodels`, then `systemone run studio` opens System One Studio (source: https://github.com/biplovgautam/LayaStudio, Apache-2.0).

## Start here

- [Every System One model](https://systemonemodels.ai/system-one-models): live table of all 30 models — maker, what each decides, parameters, context, licence, price, reported accuracy and latency
- [What is a System One model?](https://systemonemodels.ai/glossary/system-one-model): the definition, with an example
- [AI decision models](https://systemonemodels.ai/decision-models): what they are, how they differ from LLMs, classifiers and JSON mode, and every one
- [Open-source Jev alternatives](https://systemonemodels.ai/jev-alternatives): every open model you can run instead of Jev, and which speak its API
- [NoulXP, the open standard](https://systemonemodels.ai/docs/noulxp): package a System One model once, prove it matches its own code, calibrate it, run it anywhere
- [Glossary](https://systemonemodels.ai/glossary): 22 terms — typed question, choice, score, noul, calibrated confidence, ECE, decision latency…

## Models

Each link is a Markdown summary of the model page (facts, hosted API, reported evaluation, model card).

- [SAGEA: mira](https://systemonemodels.ai/sagea/mira.md): choice, score, noul, classify, route; free (open weights). Small open typed-decision model with calibrated probabilities, first-class Nepali coverage.
- [Convai Innovations: laya](https://systemonemodels.ai/convai-innovations/laya.md): choice, score, noul, classify, route; free (open weights). An open-weights System One model from Convai Innovations.
- [Fastino Labs: gliner2-5-decide](https://systemonemodels.ai/fastino-labs/gliner2-5-decide.md): choice, score, noul, classify, extract, route; hosted, price not published, or free to self-host. Fastino's open-weight decision model.
- [Supersonic Labs: julia-1](https://systemonemodels.ai/supersonic-labs/julia-1.md): choice, score, noul, classify, route; free (open weights). A 144M-parameter System One model on the multilingual mmBERT-small encoder.
- [Contrastive-LM: clm](https://systemonemodels.ai/contrastive-lm/clm.md): choice, score, noul, rank, classify, route; free (open weights). Contrastive Language Models score a state against a set of candidate actions with a contrastive objective.
- [Mapika: decider](https://systemonemodels.ai/mapika/decider.md): choice, score, noul, classify, route; free (open weights). A family of open System One models by Mark Marosi, decider-0.8b, decider-2b, decider-4b and decider-35b-a3b on Qwen3.5 bases, that softmax letter logits at an answer slot.
- [StartLux: startlux-decision](https://systemonemodels.ai/startlux/startlux-decision.md): choice, score, noul; free (open weights). StartLux's decision family in five sizes, 0.8B to 27B.
- [TokenRhythm: neohorse-jev](https://systemonemodels.ai/tokenrhythm/neohorse-jev.md): choice, score, noul, classify, route; free (open weights). TokenRhythm's prefill-only decision model for agent workflows.
- [Rizzo AI Academy: rizzo-flow](https://systemonemodels.ai/rizzo-ai-academy/rizzo-flow.md): choice, score, noul, classify, route; free (open weights). A local, Jev-compatible decision model from Rizzo AI Academy.
- [Kotoba Labs: open-jev-deberta-v3-large](https://systemonemodels.ai/kotoba-labs/open-jev-deberta-v3-large.md): choice, score, noul, classify; free (open weights). An independent Jev-shaped reproduction on DeBERTa-v3-large.
- [FrontiersMind: lumma-fev](https://systemonemodels.ai/frontiersmind/lumma-fev.md): choice, score, noul; free (open weights). FrontiersMind's decision model on a base it pre-trained from scratch.
- [Jared Palmer: kev](https://systemonemodels.ai/jared-palmer/kev.md): choice, score, noul, classify, route; free (open weights). An open family of System One models.
- [Bespoke Labs: bespoke-nimble-9b](https://systemonemodels.ai/bespoke-labs/bespoke-nimble-9b.md): choice, noul, score, classify, route; hosted, price not published, or free to self-host. An open Jev-style LoRA on Qwen3.5-9B from Bespoke Labs, trained on 2,676 contrastively curated examples to score the allowed answer tokens directly for enums, booleans and rubric levels.
- [TypeSafe AI: jev](https://systemonemodels.ai/typesafe-ai/jev.md): choice, score, noul, classify, route; $0.042 / $0 per 1m. The first System One model.
- [Together AI: tev1](https://systemonemodels.ai/together-ai/tev1.md): choice, classify, route; $0.042 / $0 per 1m, or free to self-host. Together AI's Jev-style classifier.
- [MachineFi: trio-spark-v1](https://systemonemodels.ai/machinefi/trio-spark-v1.md): choice, route, classify; $0.042 / $0 per 1m. A Situated World Model for fast judgment in agent loops.
- [Nokia: anyjev](https://systemonemodels.ai/nokia/anyjev.md): choice, score, noul, classify, route; free (open weights). A training-free layer that turns an open LLM into a decision model.
- [Cua: cua-s1-forms](https://systemonemodels.ai/cua/cua-s1-forms.md): choice; free (open weights). A 2.8 MB byte-level transformer for GUI form filling.
- [Juspay: xor](https://systemonemodels.ai/juspay/xor.md): choice, score, noul, classify, route; free (open weights). Juspay's open decision model.
- [vLLM Semantic Router: decision](https://systemonemodels.ai/vllm-semantic-router/decision.md): choice, score, noul, classify, route; free (open weights). The flagship of the vLLM Semantic Router team's Decision 1.0 family.
- [Upstage: solar-decide](https://systemonemodels.ai/upstage/solar-decide.md): choice, score, noul, classify, route; $0.1 / $0 per 1m. Upstage's structured decision model, a System One endpoint on Solar Mini 4.
- [meraGPT: state-decider-1](https://systemonemodels.ai/meragpt/state-decider-1.md): choice, score, noul, classify, route; $0.03 / $0 per 1m. meraGPT's hosted decision model (state-decider-1).
- [Respan: span-01](https://systemonemodels.ai/respan/span-01.md): noul, classify; $0.02 / $0 per 1m. Respan's behaviour-scoring model for evals, guardrails and monitoring.
- [OpenAI: decisions-api](https://systemonemodels.ai/openai/decisions-api.md): choice, classify, route; hosted, price not published. OpenAI's decision endpoint, announced at DevDay 2026.
- [Liquid AI: d1](https://systemonemodels.ai/liquid-ai/d1.md): choice, score, noul, classify, route; hosted, price not published. Liquid AI's first decision model, served only through the Liquid API.
- [Metask Lab: metask-jev](https://systemonemodels.ai/metask-lab/metask-jev.md): choice, score, noul, classify, route; free (open weights). Metask Lab's calibrated decision model in 16 languages.
- [Maisa: djev](https://systemonemodels.ai/maisa/djev.md): choice, score, noul; $0.035 / $0 per 1m, or free to self-host. A typed-decision layer for Google's DiffusionGemma, from David Villalón at Maisa AI.
- [AutoTrust AI Lab: jev-27b](https://systemonemodels.ai/autotrust-ai/jev-27b.md): choice, score, noul; free (open weights). AutoTrust's student of TypeSafe Jev 1.13.
- [Surogate (Invergent): rune](https://systemonemodels.ai/surogate/rune.md): choice, score, noul, classify, route; hosted, price not published, or free to self-host. Invergent's decision model for text and images.
- [Featherless AI: simple-jev](https://systemonemodels.ai/featherless-ai/simple-jev.md): choice, score, noul, classify, route; hosted, price not published, or free to self-host. Featherless AI's open server that turns a stock Hugging Face language model into a typed-decision endpoint.
- Plus 8 community fine-tunes: https://systemonemodels.ai/models

Facts an agent needs:

- Install the client with `pip install systemonemodels` (Python 3.10+). It provides the `systemone` command and the `systemone` Python package.
- Models are addressed as `namespace/name`; versions are immutable strings like `0.1.0`, and `latest` resolves to the newest.
- Non-interactive use: set `SYSTEMONE_TOKEN` to a personal access token (`s1_pat_…`) and skip `systemone login`, which needs a person to approve it in a browser.
- Download: `systemone pull namespace/name [--version V] [--variant FOLDER] [--dest DIR]`. Files are SHA-256 verified and cached by content.
- Publish: `systemone push DIR --repo namespace/name` for one model folder, or `systemone push DIR --all` to publish every model found under DIR; run it with `--dry-run` first. The version defaults to the next free one, and a README.md becomes the model card.
- HTTP API base: https://api.systemonemodels.ai, documented at https://systemonemodels.ai/docs/api. Errors are `{"detail", "code", "errors"}`; honour `Retry-After` on 429.

## Skill for coding agents

- [skill.md](https://systemonemodels.ai/skill.md): everything a coding agent needs to package a model for NoulXP, check it and publish it here (a Claude Code skill; any agent can follow it)

## Research

- [Calibrating decision models](https://systemonemodels.ai/blog/calibrate-classifier-confidence-temperature-scaling): six decision models on the typed-decisions benchmark; most were surer than right (Julia 1 96% sure, 72% right); `noulxp calibrate` fitted to 50 labelled requests brought confidence to accuracy with no retraining
- [What a decision costs](https://systemonemodels.ai/blog/what-a-decision-costs-gpu): decisions per second and dollars per million decisions for four decision models on an NVIDIA A40
- [Same model, different answers?](https://systemonemodels.ai/blog/same-model-different-answers-gpu-cpu): how GPU arithmetic moves a decision model's probabilities, and how NoulXP checks it
- [A decision model as an MCP tool](https://systemonemodels.ai/blog/decision-model-mcp-tool-for-ai-agents): give Claude, Cursor or any MCP client a calibrated classifier with noulxp mcp

## Docs

- [Why a different home](https://systemonemodels.ai/docs/why-a-different-home.md): System One models decide instead of write. Why they need a registry of their own rather than a corner of a hub built for language models.
- [Getting started](https://systemonemodels.ai/docs/getting-started.md): Create an account, publish a decision model, and understand what the registry stores.
- [systemone.yaml](https://systemonemodels.ai/docs/manifest.md): The manifest specification — the decision-native metadata that makes models on the registry comparable.
- [SDK and CLI](https://systemonemodels.ai/docs/sdk.md): The systemone Python package: browser login, search, pull with a shared cache, create and push.
- [CLI reference](https://systemonemodels.ai/docs/cli.md): Every systemone command and option, the environment variables it reads, exit codes, and recipes for scripts, CI and AI coding agents.
- [Inference API](https://systemonemodels.ai/docs/inference.md): Call the models the System One Engine serves from your own code: API keys, the request and the answer, plans and limits, and errors.
- [Registry API](https://systemonemodels.ai/docs/api.md): The registry API: authentication, search, downloads, publishing, device login, organizations and builds.
- [Fine-tuning with System One Studio](https://systemonemodels.ai/docs/fine-tuning.md): Train a decision model locally with System One Studio (formerly Laya Studio) and publish the result, with lineage recorded automatically.
- [Run on the System One Engine](https://systemonemodels.ai/docs/engine.md): How a decision model gets live inference on System One Models: the layouts the System One Engine runs, publishing, a live playground and analytics.
- [NoulXP: the open standard for decision models](https://systemonemodels.ai/docs/noulxp.md): NoulXP is the open standard for System One models: package a decision model once, prove it answers as its own code does, calibrate it, and run it anywhere over HTTP or MCP.
- [Make your model NoulXP compatible](https://systemonemodels.ai/docs/noulxp-compatible.md): A ten-minute guide: package a System One model so any engine runs it without your code, prove it with a conformance file, and publish it to earn the NoulXP compatible badge.

## Course

[Getting started with System One models](https://systemonemodels.ai/learn/getting-started): A free course in nine short lessons: what System One models are, how they decide, and how to call, pull, run, check, calibrate, fine-tune and publish one, with a certificate at the end. About 60 minutes.

- [Lesson 1: What is a System One model?](https://systemonemodels.ai/learn/what-is-a-system-one-model): A model that reads a situation and answers typed questions, with a probability for every possible answer, instead of writing text.
- [Lesson 2: System One models and LLMs](https://systemonemodels.ai/learn/system-one-vs-llms): When a decision model is the right tool, when a language model is, and how the two work together.
- [Lesson 3: How they decide](https://systemonemodels.ai/learn/how-they-work): What happens inside a System One model: a score for every option instead of written text, and a temperature that makes the probabilities honest.
- [Lesson 4: Your first API call](https://systemonemodels.ai/learn/first-api-call): Create an API key, ask a model two questions over HTTP and from Python, and read the answers and your usage.
- [Lesson 5: Pull a model](https://systemonemodels.ai/learn/pull-a-model): Install the CLI, find a model on the registry, and download only the files you need, checked and cached.
- [Lesson 6: Run a model on your machine](https://systemonemodels.ai/learn/run-locally): Run a System One model locally through NoulXP, serve it over HTTP, and give it to an AI agent as a tool.
- [Lesson 7: Check, measure and calibrate with NoulXP](https://systemonemodels.ai/learn/noulxp): Why run a model as a NoulXP package: prove it answers like its own code, measure it on your hardware, and fit its confidence to your own data without retraining.
- [Lesson 8: Fine-tune on your own data](https://systemonemodels.ai/learn/fine-tune): Adapt a base model to your own decisions with System One Studio, and check that it got better on examples it never saw.
- [Lesson 9: Publish your model](https://systemonemodels.ai/learn/publish): Put your model on the registry with one command, with its model card, evaluation and lineage, and learn how a model earns the NoulXP compatible badge.

## Client

- [CLI reference](https://systemonemodels.ai/docs/cli.md): every command, option and environment variable
- [systemonemodels on PyPI](https://pypi.org/project/systemonemodels/): the published package
- [Source code](https://github.com/systemonemodels/systemonemodels-sdk): Apache-2.0

## API

- [HTTP API guide](https://systemonemodels.ai/docs/api.md): authentication, search, downloads, publishing, device login

## Optional

- [All documentation in one file](https://systemonemodels.ai/llms-full.txt)
- [Browse models](https://systemonemodels.ai/models): search by capability, architecture, licence, accuracy and latency
- [Projects to build](https://systemonemodels.ai/projects): the top projects to build with System One models, each with the decision it makes, the models that suit it and step-by-step instructions
- [Built with System One models](https://systemonemodels.ai/builds): apps, agents and tools people have built with System One models
- [Compare](https://systemonemodels.ai/compare?models=a/b,c/d): up to four models side by side — accuracy, calibration, latency, size; head-to-head pages live at https://systemonemodels.ai/compare/{a}-vs-{b}, e.g. https://systemonemodels.ai/compare/jev-vs-laya
- [Hosted models](https://systemonemodels.ai/models?availability=hosted-api): commercial System One models served behind an API, with provider, price and status
- [Blog](https://systemonemodels.ai/blog): articles on decision models, Laya and Jev ([RSS](https://systemonemodels.ai/blog/rss.xml))
- [About](https://systemonemodels.ai/about): why System One models need their own home