Decision model landscape
Seven families of tools turn a text into a typed decision, proprietary and open. This page lists them product by product, with licence, deployment, output, probabilities, price and source: 39 entries, each linked to its source, checked on October 1, 2026.
Updated October 2, 2026
At a glance Hosted decision models Open decision models General LLMs with structured outputs Constrained decoding and validation Cloud classification APIs Zero-shot and few-shot encoders Guard and judge models Method and data
01 · At a glance
Only decision models, hosted or open, combine a guaranteed format, a probability per option and a single pass per question. LLMs with structured outputs guarantee the format but not the probabilities; zero-shot encoders give probabilities, on a CPU, but make no calibration claim.
The performance figures each provider publishes come from different test sets: they do not compare with each other. Our like-for-like comparisons are on the benchmarks page.
02 · Hosted decision models
Product Provider Licence Deployment Output Probabilities Price
Jev 1.13.0 TypeSafe Closed TypeSafe API, OpenRouter Label + probabilities (choice, score, noul) Calibrated according to the vendor $0.042 per million input tokens, output free d1 Liquid AI Closed Liquid API Label + probabilities Calibrated according to the vendor Free tier (d1:free); paid rates not published Decisions API OpenAI Closed API, limited preview
03 · Open decision models
Product Provider Licence Deployment Output Probabilities Price
Clef, Clef-flash Cloudflare Apache 2.0 Workers AI, weights on Hugging Face; 27B and 9B Label + probabilities, Jev-API compatible Calibrated according to the vendor Open weights; hosted on Workers AI JevK5 v0.3 allebee Apache 2.0 (weights and code) Local GPU or CPU (GGUF), /v1/systemone request shape; 4B and 9B Label + probabilities ECE published by the authors Free Kahn1 4B, Kahn1 3B Kahn1 4B Apache 2.0, 3B Qwen Research License, code MIT Local GPU or CPU, browser (3B); Jev question fields at /v1/evaluate/jev Label + probabilities Temperature-calibrated, ECE published Free Tev1-4B-experimental Together AI Being finalized (base Qwen3.5-4B, Apache 2.0) Together API, weights on Hugging Face Label (one option letter) Calibration not comprehensively evaluated (model card) open-alternative-jev (so1) ikermoel Apache 2.0 Any open LLM through Hugging Face or vLLM Label + probabilities Temperature scaling, ECE published Free Laya Convai Innovations Apache 2.0 ModernBERT 421M or mmBERT 322M; laya-serve exposes /v1/systemone Label + probabilities Calibrated according to the vendor, ECE published Free
04 · General LLMs with structured outputs
05 · Constrained decoding and validation
Product Provider Licence Deployment Output Probabilities Price
Outlines dottxt Apache 2.0 Local, any model Choice, regex, JSON Not documented per option Free XGrammar MLC Apache 2.0 Inside vLLM, SGLang, TensorRT-LLM JSON, grammars Not documented per option Free vLLM structured outputs vLLM Local Choice, regex, JSON, grammar Not documented per option Free GBNF grammars llama.cpp Local JSON, grammars Not documented per option Free Guidance guidance-ai (from Microsoft) MIT Local select, regex, JSON Not documented per option Free Instructor 567 Labs MIT Over provider APIs Validated JSON, retries Free BAML Boundary Apache 2.0 Over provider APIs Typed functions Free
06 · Cloud classification APIs
07 · Zero-shot and few-shot encoders
Product Provider Licence Deployment Output Probabilities Price
bart-large-mnli Meta MIT CPU, 0.4B, NLI Label + probabilities No calibration claim Free deberta-v3-large-zeroshot-v2.0 Moritz Laurer MIT CPU, 0.4B, NLI Label + probabilities No calibration claim Free ModernBERT-large-zeroshot-v2.0 Moritz Laurer Apache 2.0 CPU, 0.4B, NLI Label + probabilities No calibration claim Free GLiClass Knowledgator Apache 2.0 CPU, every label in one pass Label + probabilities No calibration claim Free SetFit Hugging Face Apache 2.0 CPU, few-shot fine-tuning Label + probabilities No calibration claim Free
08 · Guard and judge models
Product Provider Licence Deployment Output Probabilities Price
Moderation OpenAI Closed API Flags + scores, 13 categories Scores 0 to 1, to recalibrate when the model changes Free AI Content Safety Microsoft Azure Closed API Severity levels F0 and S0 tiers Llama Guard 4 Meta Llama 4 Community License Local, 12B Safe or unsafe + categories S1 to S14 Free ShieldGemma Google Gemma terms Local, 2B to 27B Policy score P(Yes) Free Granite Guardian 4.1 IBM Apache 2.0 Local, 8B Yes or no, your own criteria Free Qwen3Guard Qwen Apache 2.0 Local, 0.6B to 8B Safe, unsafe or controversial + categories Free gpt-oss-safeguard OpenAI Apache 2.0 Local, 20B and 120B Label + reasoning, policy given at inference Free Prometheus 2 prometheus-eval Apache 2.0 Local, 7B and 8x7B Score 1 to 5 + feedback Free Flow Judge v0.1 Flow AI Apache 2.0 Local, 3.8B Score + feedback Free Lynx Patronus AI CC BY-NC 4.0 (non-commercial) Local, 8B and 70B PASS or FAIL + reasoning Free, non-commercial HHEM-2.1-Open Vectara Apache 2.0 Local, 0.1B Consistency score Score 0 to 1 Free
09 · Method and data
Each row links, from its name, to the page it comes from: the provider's own page or repository, checked on October 1, 2026. Vendor claims (calibration, speed) are reported as claims. An empty cell means the page did not say.
Download the data: landscape.csv . Other projects listed by third-party directories (Kev, Von, NanoJev, SemIf, djev…) will join once we can check them on their own pages. Found an error or a gap? Open an issue .