Awesome Decision Models

Models that answer instead of talk: give them text and typed questions, get back calibrated probabilities in one forward pass. No generated tokens, no JSON to parse.

A curated list of decision models (also called System One models, typed decision models, or "semantic ifs"): TypeSafe's Jev, the open-source Laya, Kev, SemIf, and everything the community has built around them.

What is a decision model?

A chat LLM writes an answer token by token and your code parses it back into an if. A decision model skips the writing: you send text plus typed questions (yes/no, pick one, score) and read each answer straight off the model, with a probability, in one pass. Longer explanation with an example.

Contents

• Official
• Open models and weights
• Runtimes, ports and servers
• SDKs and libraries
• Integrations and plugins
  • Coding agents
  • Browser and computer use
  • MCP servers
• Tools and CLIs
• Apps, demos and games
• Robotics and simulation
• Training and research code
• Benchmarks and datasets
• Papers
• Articles and tutorials
• Discussions
• Related lists

Official

Jev (TypeSafe AI, hosted)

• Introducing System One Models and Jev - Launch post that introduced the System One model category.
• TypeSafe docs - API reference for Jev's typed decisions.
• typesafe-sdk-python - Official Python SDK.
• typesafe-sdk-js - Official TypeScript/JavaScript SDK.
• skills - Official agent skills for building with the System One API.
• system-one-adapter-python - Drop-in client replacement that answers System One calls with regular LLM APIs.

Laya (Convai Innovations, open source)

• Laya - Open-source, multilingual, non-autoregressive decision engine with a router across checkpoints.
• Laya docs - Guides for hooks, schema-driven decisions, Docker and LangChain.
• Laya website - Project site with demos and benchmarks.
• laya on PyPI - pip install laya, with optional server, MCP, LangChain and ONNX extras.
• Laya demo Space - Try Laya in the browser.

Open models and weights

• convaiinnovations/laya - Laya English checkpoint.
• convaiinnovations/laya-multilingual - Laya checkpoint for 100+ languages with up to 8k context.
• convaiinnovations/laya-typed-decisions - Laya fine-tuned on the typed-decisions benchmark.
• Kev - Family of Jev-like decision models on Qwen3.5/3.8 (0.8B to 27B) with training code and a Jev-compatible server.
• Kev Space - Browser demo of Kev-4B and Kev-0.8B.
• SemIf - Semantic ifs read from open models' option logits, with a WebGPU demo (formerly OpenJev).
• Tev1 - Open-weight Jev-inspired decision model fine-tuned on Qwen3.5 4B.
• Decider - System One-style model family fine-tuned from Qwen3.5, including vision variants.
• Bespoke Nimble - Data, model and recipe for an open Jev-style model.
• Von - Open-source, non-autoregressive decision model positioned as a local Jev alternative.
• Open-Jev - Open decision model (27B) returning typed probabilities without generation.
• OpenJev - Jev-compatible decision server on DiffusionGemma with image support.
• AgentJev - 0.6B decision model tuned for agent state such as diffs, traces and logs.
• Reflex - Small open decision model re-creating the System One interface on Qwen3.5.
• JevK5 - Apache-2.0 open-weight typed decision model.
• Verdict - 151M ModernBERT-based non-autoregressive decision engine.
• OpenThai-SystemOne - Thai and English System One decision model (0.8B).
• Valen - Train a Jev-like multimodal decision model with vision.
• laya-vision - Image inputs for Laya on a SmolVLM-256M backbone.
• Eikos - Single-pass typed decision models (4B and 27B) for finance.
• dev-0.4b - 0.4B bidirectional decision model on ModernBERT-large.
• Blink - System One model with an embeddable C runtime and WebAssembly support.

Runtimes, ports and servers

• laya-mlx - Native MLX runtime for Laya on Apple Silicon.
• laya-coreml - Laya on Core ML and the Apple Neural Engine.
• laya.cpp - C++ inference for Laya with CUDA, Vulkan, Core ML and CPU backends.
• receptron/laya - Run Laya from Node.js and TypeScript via ONNX Runtime.
• laya-server - Self-hosted API and web UI for Laya, compatible with the Jev API format.
• laya-apple - Heterogeneous Laya runtime using MLX GPU and the Neural Engine.
• laya-mps - Run typed decisions on a Mac via PyTorch MPS with low RAM use.
• layaForWeb - Laya running in the browser.
• open-jev-laya - Browser-only Laya multilingual demos on Transformers.js.
• laya-goish - Laya GGUF inference and HTTP server.
• lev - Jolt implementation of the Laya decision engine.
• Arbiter - Serve Laya or custom decision models on NVIDIA or Apple Silicon behind a Jev-compatible API.
• Ollaya - Pull and serve open decision models (Laya, Decider, NLI) behind a TypeSafe-compatible API.
• EdgeJev - CPU-only ONNX + INT8 inference for Laya, Kev and similar models.
• openjev-sglang - Jev API server on SGLang with prefill-only open models.
• djev - Structured decisions on DiffusionGemma, from a vLLM example server.
• vllm-jev - Native vLLM serving for Jev-style decision models.
• JevFire - Parallel decisions over one context for CUDA LLMs with a vLLM API.
• jevmlx - Parallel constrained decisions for any MLX model.
• JEV-CPU - Run SemIf decisions on CPU with a web UI.
• jeff - Self-hosted Jev replacement powered by GliFormer.
• JevCache - Decision cache that memoizes repeated calls to Jev-class models.

SDKs and libraries

• AnyJev - Turn any LLM into a Jev-style decision model without training.
• simple-jev - Classification and scoring endpoint over open Hugging Face models via next-token logits.
• LLM2Jev - Structured decisions from local LMs using prefill only, for text and images.
• rizzo-flow - Local typed decisions from an LLM without generating tokens.
• open-jev - Browser-focused TypeScript library for typed decisions.
• system-one - Provider-neutral System One runtime for TypeScript.
• go-jev - Go SDK and CLI for Jev.
• ruby_decision_model - Ruby client for decision models such as Jev.
• laya_ex - Elixir library wrapping Laya.
• jev-foundation-models - Swift bridge between Jev and Apple's Foundation Models framework.
• advocaat - Small type-safe client for asking questions about your data with Jev.
• semantic-if - SemIf method on vLLM as a Python library, CLI, REST API and MCP server.

Integrations and plugins

Coding agents

• fast-jev-compaction - Claude Code plugin that scores tool calls with Jev to drop stale context instead of summarizing.
• jev-pruner - Claude Code plugin that trims long Bash output with Jev.
• winnow - Claude Code context sieve that judges each tool result with a System One model.
• hermes-jev-skills - Jev-powered routing, memory, compaction and skill selection for Hermes, Claude Code and Codex.
• jev-router - Route Claude Code tasks to the cheapest capable model.
• jev-codex-router - Per-turn model and reasoning-effort routing for Codex.
• jev-rules - Select which rules apply to each prompt so Claude only sees relevant ones.
• skillranker - Rust CLI that ranks agent skills for the next step, with Claude Code hooks.
• pi-jev - Jev decision layer and tool-call gate for the Pi coding agent.
• jev-review - Staged code-review workflow with a local dashboard.
• building-with-jev-skill - Agent skill for writing and improving programs that call Jev.
• keel - Local-first macOS coding workspace with local Laya decisions.
• laya-drift - OpenCode plugin that monitors agent drift with Laya.
• jevals - Agent evals and guardrails as decision calls, runnable locally with Kev or Laya.

Browser and computer use

• jev-ultrafast - Browser agent where Jev picks each operation and element from an indexed action space.
• laya-ultrafast - jev-ultrafast running on local Laya.
• laya-browser-agent - Local Playwright browser agent driven by Laya.
• jev-browser-use - Browser operations where Jev clicks and Codex plans and verifies.
• fastbrowse - Browser agent with page-cited answers; Jev picks actions, an LLM plans.
• jev-browser - Browser automation library, CLI and MCP server where an LLM plans and Jev decides.
• jev-voice-browser - Control a real browser by voice.
• FluidUse - On-device computer use on macOS with Laya via the Accessibility API.
• jev-use - macOS computer-use harness that reads the screen via Accessibility.
• Jev-cu - Hands Codex Computer Use's next-click choice to Jev.
• mobile-jev - Android phone agent with Jev making the decisions.
• third-hand - Computer-use assistant built on decision models.
• system1-agents - Jev, Laya and Cua-S1 as the brain for browser, computer, game and robot agents.
• unclutter - Browser extension that removes page clutter with reusable rules.

MCP servers

• jev-mcp - Jev typed judgments as MCP tools.
• system-one-connector - MCP connector for Jev and open-weight models such as Laya.
• Jevbridge - ACP and MCP adapter bridging Jev with Codex, Claude, Grok and OpenCode.
• jev-sift - Agent plugin and MCP tool for batch text classification.

Tools and CLIs

• gutcheck - grep for meaning: a single-binary CLI that runs Laya locally, with a live explorer, named questions and a log cache.
• jgrep - grep where the pattern is a description.
• jev-semgrep - Grep by meaning across languages with AND/OR/NOT.
• jegrep - Find code by describing it.
• jevgrep - Semantic code search for coding agents via CLI or MCP.
• perch - Semantic code linting.
• semdecide - Typed semantic decisions for Unix pipelines and CI.
• webctl - Web search CLI for agents.
• jev-shell-history - zsh history autosuggestions ranked by Jev.
• pg-jev - PostgreSQL extension to ask tables questions in plain language.
• docjev - Fast document classifier and splitter.
• jev-curate - Rust dataset sifter for filtering pretraining and synthetic data.
• jevcal - Calibrate thresholds and check drift for typed decision models.
• jeview - Local visualizer for every Jev call your code makes.
• HA-Jev - Home Assistant integration exposing typed answers as sensors.
• hono-jev-router - Route HTTP requests by meaning in Hono.

Apps, demos and games

• laya-playground - Website, games, a benchmark and an agent skill for local Laya.
• laya-vs-jev - Local Laya and hosted Jev playing T-Rex side by side.
• laya-vs-jev-arena - Laya and Jev racing in Snake and fighting in an arena.
• jev-experiments - Collection of latency-focused Jev demos.
• shapeshift - A text box that morphs into the right UI as you type.
• jev-chat - Tool-calling chatbot built with Jev and no LLM.
• jevmail - Gmail inbox triage.
• jev-search - Web search with Jev for source selection and ranking.
• tax-doc-classifier - Tax document page classifier.
• jev-chat-jarvis - Phone chat copilot that reads the screen and suggests replies (Chinese).
• jev-trader - One trade decision per Monad block.
• typesafe-mario - Agent that plays Super Mario Bros. from emulator state.
• minecraft-agent - Minecraft planner with a Jev controller.
• jevpilot - Three.js driving simulator with a Jev autopilot.

Robotics and simulation

• quackd - CLI to command robots, with a decision model for multiple-choice steps.
• embodied-jev - MuJoCo robot decision workbench.
• jev-drone - Camera-only autonomous drone in MuJoCo with a decision model in the loop.
• jev-libero - Fine-grained robot control on LIBERO tasks.

Training and research code

• NanoJev - Nano replica of Jev with an end-to-end training pipeline.
• Jevlike - Train a small model that chooses among a changing list of options.
• jev-forge - Training and inference stack for Jev-style models.
• autojev - Full-weight Qwen decision model with training code and a playground.
• tinyjev - Tiny choice/score/noul model in MLX or PyTorch.
• mini-jev - Preregistered experiment reading option logits from a frozen Qwen3-4B.
• jev-on-a-laptop - Study of Jev-style decisions on stock 1.5B-8B models on Apple Silicon.
• sarvam-jev - Generation-free typed decisions on Indic LLMs.
• JevHarness - LLM-authored task-specific Jev harnesses with GEPA evolution.
• jev-align - Build calibrated AI functions from human feedback with Jev and GEPA.

Benchmarks and datasets

• typed-decisions - Typed-decisions benchmark dataset used in several model cards, including Laya's.
• kev-suites - Frozen evaluation suites from the Kev project.
• laya-jev-benchmark - Head-to-head Laya vs Jev benchmark data.
• JevBench - Benchmark for Jev-class models on accuracy, cost, speed and reliability.
• jev-benchmarks - Calibration, selective risk and latency evaluation for typed decision models.
• jev_and_laya_benchmarking - Scripts comparing Jev and Laya.
• JevPokerBench - Texas Hold'em benchmark and leaderboard for decision models.
• AIM-Decision - Independent comparison of Jev, Kev and LLMs.

Papers

• SalesRLAgent (arXiv:2503.23303) - Earlier reinforcement-learning paper by Laya's author, cited as prior work.
• Confidence-Aware Routing for LLM Reliability Enhancement (arXiv:2510.01237) - Pre-generation routing paper by Laya's author.
• Calibrated Decisions at Scale (arXiv:2609.24052) - Converting police crash narratives into probabilistic variables with Jev.
• Calibrated Decision Models for Autonomous Penetration-Testing Harnesses (arXiv:2609.28940) - Jev and Laya as decision layers for pentest agents.

Articles and tutorials

• Jev's Architecture Unmasked - Reverse-engineered architecture of Jev that Kev builds on.
• I Built Non-Autoregressive Decision Models a Year Ago - Laya's author on the background of the project.
• A deep dive into Jev - Walkthrough of the System One model.
• How to Use Jev - Practical guide.
• A Coding Guide to TypeSafe AI Jev - Typed decisions, confidence and fan-out in code.
• Building a harness with Jev - LangChain's guide to the System One model.
• Jev: TypeSafe's System One Model - DataCamp explainer.
• Laya: A 421M Local Decision Model - How Laya works under the hood.
• Laya: The Open-Source Jev Alternative, Benchmarked - Laya benchmarked against Jev.
• Jev vs Laya - Choosing between closed and open System One models.
• Jev vs djev vs Laya vs OpenJev vs SemIf - Comparison of open decision models.
• Open-Weights Jev Alternatives - Which open decision model to ship.
• Best Open Source Jev Alternatives - Self-hosting decision models.
• Here are 6 Clones of Jev in 2 days - Latent Space roundup of the first open clones.
• TypeSafe AI Releases Jev - Launch news coverage.
• TypeSafe AI's Jev offers an alternative to LLMs - Tom's Hardware coverage.

Discussions

• Kev on Hacker News - Discussion of Kev.
• Laya author on Hacker News - Discussion of the Laya background post.

Related lists

• awesome-jev (yibie) - Projects, integrations and discussions built on Jev.
• awesome-jev (heyjunpenn) - Large catalog of open-source projects built with Jev.
• awesome-jev-tools - Tools built for Jev.
• awesome-jev-zh - Chinese-language list for Jev and System One.

Contributing

Contributions welcome! Read the contribution guidelines first.