Decision model runtimes and harnesses: vLLM, MLX, APIs

Serving and orchestration for decision models: vLLM pooling, MLX on Apple Silicon, and hosted APIs that answer typed questions without trained weights.

These are NOT models. They extract decisions from existing LLMs via constrained decoding, logit readout, or wrapper APIs. No trained decision weights — the decision capability comes from inference-time techniques on frozen base models.

Looking for trained decision models? See the Model catalog.

Aleria AI

Aleria Judge (MORJ)

Method

Aleria AI's open method that turns stock Gemma 4 31B into a typed decision server (formerly MORJ). Board entry MORJ 1.0.1: Full 60.96 on Decision Index 0.3.2, tied with Jev 1.13.

Why not a model

Training-free: it reads stock Gemma-4-31B next-token probabilities over rotated option orders. No decision weights of its own, so no model page (precedent: simple-jev, AnyJev).

Base LLMs
google/gemma-4-31B-it (stock), nvidia/Gemma-4-31B-IT-NVFP4 (TensorRT-LLM, text only)
Types
choice · score · boolean
Stars
~17
Full details

Nokia Applied Research

AnyJev

Method

Nokia’s open library that turns any causal LLM into a Jev-style decision model without training, with option-order debiasing.

Why not a model

Training-free: reads typed answers from any LLM’s next-token logits, averaged over option rotations (L0). No decision weights of its own.

Base LLMs
any causal LLM (Transformers, vLLM, SGLang, llama.cpp), Qwen3-8B (examples)
Types
choice · score · boolean
Stars
~1,167
Full details

bandr.ai (bandr-ai)

Bandits · Your own Jev recipe

MethodResearch

Recipe in the Bandits toolkit that LoRA-tunes Qwen3.5-4B-Base into a per-step agent-trace judge on your own labels. No checkpoint, no license, no inference server.

Why not a model

Ships a training recipe, not a model: LoRA on Qwen3.5-4B-Base, local training UI, scorecard vs Jev via /v1/systemone. No released weights, no license.

Base LLMs
Qwen/Qwen3.5-4B-Base
Types
choice
Stars
~28
Full details

classifier.dev

classifier.dev

Wrapper

HTTP/CLI/MCP wrapper over hosted Jev with smart escalation to reasoning LLMs for uncertain cases.

Why not a model

Hosted wrapper over Jev API — not a separate model, just a convenience layer.

Base LLMs
jev-1.13
Types
choice · score · boolean
Full details

Cua AI

CUA-S1-FORMS

SpecialistResearch

Lightweight specialist (~706k params) for scoring form field actions: use, check, click, skip.

Why not a model

Narrow-task specialist for form field actions — not a general-purpose System One decision model.

Base LLMs
custom-706k
Types
choice
Stars
~706
Full details

vLLM Semantic Router

Decision 3.0 Free API

Runtime

Free hosted System One API for six existing Decision 3.0 checkpoints, with keys available without an account.

Why not a model

Hosted API serves the existing Decision 3.0 checkpoints; it publishes no separate decision-trained weights of its own.

Base LLMs
vLLM-SR d3 family (six sizes, 0.6B–27B)
Types
choice · score · noul
Full details

Decision Models (Florian S)

Decision Models / REAX

Runtime

Hosted Jev-compatible API serving open decision checkpoints through REAX, with Peer-to-Peer and EU tiers.

Why not a model

Hosted inference runtime over existing open checkpoints; REAX is a serving engine, and the service has no separately published decision-trained weights.

Base LLMs
vLLM-SR d3, Blink v0.3, Quyet-1.0-Large…
Types
choice · score · noul
Full details

FrontiersMindAI

DecisionFormer

Runtime

Transformers-style library to fine-tune and serve calibrated decision models on any HF backbone: choice / noul / score in one forward pass, /v1/systemone server, QLoRA, vLLM backend.

Why not a model

A library to fine-tune, load and serve decision heads on any Hugging Face model. It ships no checkpoint of its own; quality comes from the backbone and data you train.

Base LLMs
any Hugging Face backbone, vllm-sr/d3-edge (README example), Qwen3.5 (Gated DeltaNet kernels)
Types
choice · noul · score
Stars
~20
Full details

wcole3 (independent)

jev-brain

Wrapper

One GPU, one model, two APIs: a vLLM chat server plus a Jev-compatible /v1/systemone gateway over the same unmodified model.

Why not a model

A gateway that reads option-letter logprobs from an unmodified chat model served by vLLM. The weights are unchanged; no decision training of its own.

Base LLMs
Gemma-4-31B (Decision Index run), any vLLM-served chat model
Types
choice · score · noul
Full details

mizorewww (independent)

Laya-MLX

Runtime

Native MLX runtime for ConvAI's Laya typed decision models on Apple Silicon, without PyTorch or a cloud API. Version 0.5.0 follows upstream Laya v0.4.12.

Why not a model

Native MLX port of ConvAI's Laya: converted FP16 checkpoints and an inference engine for Apple Silicon. No decision training of its own.

Base LLMs
Types
choice · score · noul
Full details

notnotsamuel

LFM2.5-350M-RLCD

MethodResearch

Research method for parallel constrained scoring on frozen LiquidAI LFM2.5-350M weights.

Why not a model

Parallel constrained scoring on unchanged LiquidAI weights — name 'RLCD' ≠ TypeSafe RLCD training method.

Base LLMs
LiquidAI/LFM2.5-350M
Types
choice
Full details

zhengxuyu

litjev

Method

Lightweight method that wraps Qwen checkpoints to serve TypeSafe-compatible decisions locally.

Why not a model

Wraps off-the-shelf Qwen checkpoints as Jev-schema decisions — probabilities are uncalibrated by default.

Base LLMs
Qwen
Types
choice · score · boolean
Full details

Google's on-device Decision Maker task scores choice, boolean and rubric questions on EmbeddingGemma, Laya or Gemma 4, including Jev-shaped JSON.

Why not a model

On-device task over EmbeddingGemma, Laya, GLiNER2.5-Decide and Gemma 4. It does not train or ship its own decision weights.

Base LLMs
EmbeddingGemma 2, EmbeddingGemma 1, Laya…
Types
choice · score · boolean
Full details

Ollama

Ollama

Runtime

Ollama v0.35+ adds a native, TypeSafe-compatible /v1/systemone endpoint and three decision models: nimble, tev1 and tev1:0.8b. Local only for now.

Why not a model

Inference server. Since v0.35 it serves decision models trained by others (Nimble, Tev1) on a local /v1/systemone endpoint; it has no decision weights of its own.

Base LLMs
Bespoke-Nimble-9B, Tev1-4B-experimental, Tev1-0.8B-experimental
Types
choice · noul · score
Stars
~182,000
Full details

ollaya-dev

Ollaya

Runtime

Ollama for decision models: pull and serve open decision models locally behind a TypeSafe-compatible API.

Why not a model

Serves existing open decision models locally (Laya, Decider, NLI, GLiClass, Kev, Von); no trained weights of its own — it's distribution and inference infra.

Base LLMs
Types
choice · score · boolean
Stars
~935
Full details

ikermoel (Iker Moel Tacher)

Open Alternative to Jev

Runtime

Python library (import so1) that packs several typed questions into one forward pass of a stock open LLM and reads option-letter logits. HF Transformers and vLLM backends.

Why not a model

Reads option-letter logits from stock open LLMs (Qwen) in one packed forward pass. No decision training; calibration is an optional temperature you fit yourself.

Base LLMs
Qwen3.6-27B, Qwen3.5-4B, Qwen3.5-2B…
Types
choice · boolean · score
Stars
~54
Full details

OpenAI API (GA since 22 Oct 2026) that answers predicate, choice and score questions with gpt-6-luna at $0.10/M input. Not the same product as OpenRouter's or Perplexity's.

Why not a model

Documented as gpt-6-luna behind a typed endpoint; no decision-trained weights, head or model card. GA since 22 Oct 2026, same model.

Base LLMs
GPT-6 Luna (gpt-6-luna)
Types
choice · score · boolean
Full details

deepanwadhwa (independent)

OpenDecision

Wrapper

Open-source, Jev-style decision engine built on a frozen ModernBERT-large zero-shot NLI model: Choice / Noul / Score / Relation, a document API and a /v1/systemone endpoint.

Why not a model

Wraps a stock, frozen zero-shot NLI classifier (ModernBERT-large) behind a Jev-style API. No decision-specific training; the author says the scores are uncalibrated.

Base LLMs
MoritzLaurer/ModernBERT-large-zeroshot-v2.0 (default), MoritzLaurer/deberta-v3-large-zeroshot-v2.0
Types
choice · noul · score
Stars
~57
Full details

OpenDecisions (unikcc, OneJev maintainer)

OpenDecisions

Runtime

One server for decision models: runs OneJev locally (PyTorch or llama.cpp) behind /v1/decisions and /v1/systemone, and can route low-confidence questions to hosted Jev.

Why not a model

Server and router. Serves the OneJev weights (catalogued separately) and can fall back to hosted Jev; it has no decision weights of its own.

Base LLMs
OneJev-4B (default), OneJev 0.8B / 9B / 27B, Jev (hosted, via TypeSafe or OpenRouter)
Types
choice · noul · score
0
Full details

ekzhang

openjev-sglang

Runtime

SGLang-based runtime providing TypeSafe-compatible /v1/systemone endpoint over Qwen3.6-35B-A3B.

Why not a model

TypeSafe-compatible endpoint over Qwen via SGLang prefill/logprob hack — not a trained decision model.

Base LLMs
Qwen3.6-35B-A3B
Types
choice · score · boolean
Full details

Parallel Web Systems

Parallel Decisions API

Wrapper

Parallel's web-grounded decisions API (launched 27 Oct 2026): typed yes/no, choice and score answers with sources, $1 per 1,000 calls, 5,000 free requests.

Why not a model

Hosted service that retrieves live web evidence and answers typed questions; Parallel publishes no decision model, weights or card behind it.

Base LLMs
undisclosed
Types
boolean · choice · score
Full details

alperiox

Prosodia

SpecialistResearch

Audio-native Jev-shaped decisions without ASR text decode — processes audio directly.

Why not a model

Audio modality specialist — not a text-based System One peer; different input domain entirely.

Base LLMs
whisper-encoder
Types
choice · score · noul
Full details

Decision Models (decisionmodels.io)

s1-llm-auto-router

Specialist

Small MIT-licensed router (mmBERT-small, ONNX) that answers seven fixed questions about a request on CPU in ~26 ms, so you can pick an LLM. Also on the Decision Models API.

Why not a model

Single-task LLM-routing specialist: it answers only seven fixed questions with fixed option sets, so it does not compete with general System One models.

Base LLMs
mmBERT-small
Types
choice · score · noul
Full details

TheoLeeCJ

SemIf

Runtime

Browser-first runtime that extracts typed decisions from frozen open LLMs via WebGPU/wllama.

Why not a model

Reads typed option logits from frozen open LMs — no trained decision weights, just inference-time logit extraction.

Base LLMs
Qwen3.5-4B, MiniCPM5-2B, MiniCPM5-4B…
Types
choice · score · boolean
Stars
~1,900
Full details

Featherless AI

simple-jev

Runtime

Featherless stack that converts any open LLM into a typed decision classifier via next-token logits.

Why not a model

Turns any compatible open LM into a typed classifier via next-token logits — no fine-tuning required.

Base LLMs
any-open-lm
Types
choice · score · noul
Full details

Snowflake

Snowflake Decision

Wrapper

Snowflake’s decision model behind AI_COMPLETE (model snowflake-decision), in private preview: choice, score and true/false over table rows in SQL.

Why not a model

Private-preview Cortex model; base model, weights and training are undisclosed. Re-check if Snowflake documents a decision-trained model or opens it up.

Base LLMs
undisclosed
Types
choice · score · boolean
Full details

Upstage

Solar Decide

Wrapper

Upstage's beta System One endpoint (POST /v1/systemone) and OpenRouter listing that reads decisions from Solar Mini 4's label logits. Up to 26 options per question.

Why not a model

Upstage serves it on Solar Mini 4 and reads the decision from the logits of single-token labels A–Z. No decision-specific weights, head or model card documented, and no independent benchmark.

Base LLMs
Solar Mini 4
Types
choice · score · noul
Full details

StartLux Labs

StartLux-Prism

Runtime

Fast NVIDIA engine for StartLux-Decision GGUF files behind /v1/systemone: int8 and 4-bit tensor cores, bit-identical answers, author claims up to 5.3x lower latency than vLLM FP8.

Why not a model

Inference engine for the existing StartLux-Decision GGUF files on NVIDIA GPUs. Prompts, read-outs and decision rules come from the model; only the forward pass is new.

Base LLMs
StartLux-Decision v1 (4B, 9B, 27B), StartLux-Decision v1.1 AutoThink (4B, 9B, 27B)
Types
choice · noul · score
Stars
~57
Full details

Hugging Face

transformers serve

Runtime

Transformers main (PR 50054, merged 9 Nov 2026) adds DecisionPipeline for choice, score and noul. Not in the v5.22.2 tag. The October /v1/systemone endpoint remains.

Why not a model

Reference inference server. Since v5.20.0 `transformers serve` exposes /v1/systemone for any autoregressive model; it has no decision weights of its own.

Base LLMs
any autoregressive model
Types
choice · noul · score
Stars
~168,000
Full details

TypeLLM

TypeLLM

Runtime

SGLang-based runtime for type-safe generation on autoregressive LLMs, with thinking mode and image input. Open source, and since 29 Sep 2026 also a hosted playground and API.

Why not a model

Constrained decoding layer over existing autoregressive LLMs — does not change architecture or weights.

Base LLMs
Qwen3.8-27B, Qwen3.5-0.8B, Qwen3.5-4B…
Types
choice · boolean · integer · number · string
Stars
~885
Full details

vLLM project

vLLM (upstream)

Runtime

vLLM v0.32.0 (18 Oct 2026) serves /v1/systemone by default: choice, noul and score questions read from first-token log-probabilities. Separate from the third-party vllm-jev project.

Why not a model

Inference engine. Since v0.32.0 it serves /v1/systemone for models trained by others; it ships no decision weights and reads answers from first-token log-probabilities.

Base LLMs
Clef-Omni, Clef-Flash, vllm-sr d3 family…
Types
choice · noul · score
Stars
~94,000
Full details

mode-io

vLLM Jev

Runtime

Native vLLM (Linux) and Apple Silicon serving for open Jev-style decision checkpoints: Choice/Noul/Score over HTTP /v1/systemone, including Valen multimodal.

Why not a model

Serving engine for existing Jev-style checkpoints (vLLM on Linux, MLX/MPS on Mac). No decision training — it selects a protocol and runs inference.

Base LLMs
ZefanCai/Open-Jev-2B, OpenJev-0.6B, Tiny-Jev…
Types
choice · boolean · score
Stars
~94
Full details