There is no drop-in replacement for Jev. What exists splits three ways: open reproductions of its API built on frozen or fine-tuned open models, zero-shot classifiers and NER models that decide over labels you supply, and structured-output libraries that make any LLM return a typed answer. None of them is trained the way Jev is.
Choosing one? Read the Jev alternatives guide, which compares hosted and open-source options on benchmarks, hardware and licence.
Jev is the first row, since it is what the rest are copying. Star counts were read from the GitHub API on 30 September 2026 and move daily. Every project below was published within a month of Jev's launch on 15 September 2026, so treat all of it as young code rather than as a settled field. For a verdict on the hosted and open options, read theJev alternatives guide.
| Name | Type | Base model / params | Weights and licence | Runs on | Primitives | Calibration claim | Stars |
|---|---|---|---|---|---|---|---|
| Jev 1.13 (TypeSafe AI) | Hosted System One model | Not published | Closed weights, API only | TypeSafe's API | Choice, Score, Noul | TypeSafe says RLCD training gives "epistemically honest probabilities" | n/a |
| Laya (Convai Innovations) | Decision head trained on a frozen encoder | ModernBERT-large, 421M (English); mmBERT-base, 322M (multilingual) | Open, Apache-2.0 | Tesla T4 GPU tested, 32.8-39.5ms; CPU, 193-464ms | Choice, Score, Noul | ECE 0.081 after temperature fitting, from 0.466 as shipped (model card) | 28,759 |
| Kev (Jared Palmer) | LoRA adapter and pointer head on a frozen base | Qwen3.5-Base at 0.8B, 4B and 9B; post-trained Qwen3.8-27B at 27B | Open, Apache-2.0, training code and eval data included | CUDA, ROCm or Apple Silicon (MLX); Kev-4B 41.5-145ms model time on an L40S (README) | Choice, Score, Noul | Temperature fitted per checkpoint; Kev-9B Brier 0.237 on its new-source test set (README) | 7,948 |
| Von (wfzyx) | Encoder that scores each option in one pass | ModernBERT, 395M | Open, Apache-2.0 | CUDA, ROCm, Apple MPS, Intel GPUs via OpenVINO, or CPU; about 23ms on an A10G GPU (README) | Choice, Score, Noul | Temperature map fitted on public JevBench items; sealed ECE 0.107, which the model card says may not carry to other data | 776 |
| CLM (Contrastive-LM) | Contrastive head that scores a state against each action | Qwen3-8B encoder plus a 75 MB head | Open, Apache-2.0 | Linux with an NVIDIA GPU; 28ms for a new state on an RTX 4090 (README) | Choice, Score, Noul | None claimed; an optional temperature divides the logits | 2,540 |
| Tev1 (Together AI) | LoRA fine-tune that picks one option letter | Qwen3.5-4B | Open weights with no licence declared on Hugging Face, MIT training code | Together's serverless API at $0.042 per million input tokens, or your own GPU | Choice | Disclaimed: "Logprobs are model preferences, not calibrated confidence" | 173 |
| Solar Decide (Upstage) | Hosted System One model, in beta | Solar Mini 4, 35B mixture-of-experts with 3B active | Closed weights, API only | Upstage's API at $0.10 per million input tokens, or OpenRouter | Choice, Score, Noul | OpenRouter's listing calls the probabilities calibrated; no figure published | n/a |
| Span-01 (Respan) | Hosted behavior classifier for agent traces | Not published | Closed weights, API only | Respan's API at $0.02 per million input tokens, or OpenRouter; Span-01 Lite is free | Present, absent or not observable per behavior (closest to Noul) | No calibration figure published | n/a |
| d1 (Liquid AI) | Hosted System One model, plus two open-weight versions | Not published for the hosted d1; LFM2.5-VL-3B (d1-3B); LFM2.5-Encoder-350M (d1-omni-600M) | Hosted d1 closed; d1-3B and d1-omni-600M under the LFM Open License v1.0 | Liquid's API as d1:free or the paid d1, price not published; OpenRouter at $0.04 per million input tokens; open versions on your own hardware | Choice, Score, Noul | None for the hosted d1; d1-3B 0.047 on the Jev Decision Index (multimodalart's community board on Hugging Face), with Liquid running the public part | n/a |
| Mercury Decide (Inception) | Hosted System One model | Not published | Closed weights, API only | Inception's API or OpenRouter at $0.02 per million input tokens; free tier on OpenRouter | Choice, Score, Noul | No calibration figure published | n/a |
| Clef (Cloudflare) | Hosted System One model with open weights, reads images | Qwen3.8-27B (Clef); Qwen3.5-9B (Clef-flash) | Open, Apache-2.0 | Cloudflare Workers AI or OpenRouter at $0.24 per million input tokens (Clef) or $0.038 (Clef-flash), or your own GPU | Choice, Score, Noul | Trained with a Brier loss and Cloudflare's own RLCD stage; calibration error 0.040 (Clef) and 0.025 (Clef-flash) on the Jev Decision Index, against 0.074 for Jev | n/a |
| pplx-decider (Perplexity) | Hosted System One model with open weights, reads images | Qwen3.8-27B | Open, Apache-2.0 | Perplexity Decisions API at $0.02 per million input tokens, or your own GPU | Choice, Score, Noul | Calibration error 0.071 for v1.1 on the Jev Decision Index, against 0.074 for Jev | n/a |
| Kai (Hanzo AI) | Hosted System One model | mmBERT-base encoder initialised from Laya, per Hanzo | Closed weights, API only | Hanzo's API at $0.021 per million input tokens | Choice, Score, Noul | Temperature-scaled on held-out rows; calibration error 0.093 against 0.139 for Jev on Hanzo's own 12 suites, trained on their train splits | n/a |
| Celeris-1 Decision (Celeris) | Hosted System One model, reads images | Not published | Closed weights, API only | Celeris's API at $0.04 per million input tokens | Noul, Choice, Score | Brier 0.272 against 0.380 for Jev 1.13.0 on jev-bench, computed by Celeris | n/a |
| Microsoft-Decision-1 (Microsoft) | Hosted System One model | Qwen3.5-9B, post-trained by Microsoft | Closed weights, API only | Microsoft Foundry and OpenRouter at $0.042 per million input tokens | Choice, Score, Noul | Calibration score 92.2 of 100 over 36 benchmarks, second to Quyet-1.0-Large, in Microsoft's own run; Jev not scored | n/a |
| Clef-omni (Cloudflare) | Hosted System One model with open weights, reads images, audio and video | Qwen3-Omni-30B-A3B, mixture-of-experts with 3B active | Open, Apache-2.0 | Cloudflare Workers AI or OpenRouter at $0.15 per million input tokens, or your own GPU | Choice, Score, Noul | Trained with a Brier loss, per Cloudflare; no calibration figure published and not on the Jev Decision Index | n/a |
| Drex (Nace.AI) | Hosted System One model with restricted open weights | MiMo-V2.6-Distill-Qwen-9B, 9B, post-trained by Nace | Nace.AI Open RAIL-M, commercial use limited to small organisations | Nace.AI's API at $0.05 per million input tokens, OpenRouter at $0.04, or your own GPU | Choice, Score, Noul | No calibration figure published; not on the Jev Decision Index | 3 |
| NanoJev (TianyuCodings) | Trained from scratch, 0.6B | 0.6B, own architecture | Open, MIT, weights and dataset included | CUDA GPU | Choice, Score, boolean | Calibration split in the training pipeline, no metric reported | 2,440 |
| SemIf, formerly openjev (TheoLeeCJ) | Frozen-model logit reader | Qwen3.5-4B, per apidog | Open, MIT | Consumer GPU, RTX 3090 class | Choice | None claimed | 4,583 |
| jevlike (vinnylarouge) | Trained from scratch, own architecture | Byte embeddings with option attention | Open, MIT | CPU, Apple Silicon or CUDA | Choice | None claimed | 1,336 |
| openjev-sglang (ekzhang) | Frozen-model server, TypeSafe's wire format | Qwen3.6-35B-A3B on SGLang | Open code, no LICENSE file detected | B200-class GPU, deploys to Modal and scales to zero | Choice, Score, Noul | Disclaimed: "not calibrated estimates of correctness" | 334 |
| Decider (Mapika) | Fine-tuned small model | Qwen3.5-2B-Base | Open, Apache-2.0 | GPU, local inference of a 2B model | Choice, Score, Noul | Claims calibrated probabilities, no measurement published | 979 |
| OpenDecision (deepanwadhwa) | Zero-shot NLI classifier | MoritzLaurer/ModernBERT-large-zeroshot-v2.0, ~400M | Open, Apache-2.0 | No GPU stated as required, downloads once from Hugging Face | Choice, Score, Noul | Disclaimed: confidence measures concentration, not correctness | 58 |
| mini-jev (r-ms) | Frozen-model logit reader | Qwen3-4B-Instruct-2507 | Open, MIT | About 8.5 GB of memory, Apple Silicon or CUDA | Choice, Noul | Disclaimed: "a ranking with a confidence gap, not calibrated probabilities" | 58 |
| jev-on-a-laptop (rorshopping) | Frozen-model study with a demo Space | Stock 1.5B to 8B models | Open, no SPDX licence detected | Apple Silicon laptop | Choice | None claimed | 25 |
| Parallel Constrained Decoding (drinkmoonshine) | One-pass schema scorer, Hugging Face Space | Qwen2.5-1B-RLCD, called Qwen2.5-1.5B-Instruct-4bit by apidog | Open, Apache-2.0 | Apple Silicon under MLX | Choice, Score | RLCD is in the name, but the README documents no training of any kind | n/a |
No. Jev is a closed model served from one hosted endpoint, and nothing about how it works has been released.
What is public: the HTTP API, the documentation, the SDKs, the prices and the jaggedness page listing the tasks the model is bad at. What is not public: the weights, the parameter count, the base architecture and the RLCD recipe. TypeSafe's docs describe Jev only as an endpoint, and say "the same weights serve every account," with behaviour shaped through the request rather than through per-account fine-tuning. The architecture guide collects the whole public record, which runs to five statements.
No. There is no weight download, no self-hosted build and no announced VPC option as of 30 September 2026, so running a System One model on your own hardware means running a different model.
How much hardware that takes depends on how close to Jev you want to get. An encoder-sized classification model needs no GPU at all: OpenDecision pulls a 400M ModernBERT checkpoint from Hugging Face on first use, and Laya reports 193ms to 464ms per question on CPU against 32.8ms to 39.5ms on a Tesla T4. Apple Silicon covers the MLX projects, among them jevmlx, jev-visual and mini-jev, which wants about 8.5 GB of memory. Kev runs there too through MLX, and its 4B and 9B models fit a 32 GB Mac. A single consumer GPU runs SemIf on an RTX 3090, NanoJev on CUDA and CLM on Linux with an NVIDIA card. Only openjev-sglang asks for a datacenter card, since it serves a 35B model on SGLang and expects a B200.
You send a state and a list of typed questions, and one forward pass comes back with a probability for every option you supplied. Several of them go further and speak TypeSafe's own wire format, so existing client code keeps working: openjev-sglang implements POST /v1/systemone along with /v1/models and /v1/limits, while Decider, litjev, Kev, CLM and Von serve the same schema, which means repointing TYPESAFE_BASE_URL is the whole migration.
What they do not reproduce is the training. RLCD has never been described in enough detail to copy, so a frozen Qwen reading its own option logits approximates the interface while the method stays out of reach. Most of the authors say as much in their own READMEs. openjev-sglang states that its probabilities "are not calibrated estimates of correctness," and jevfire that "relative label probabilities are not probabilities that an answer is correct." The nearest thing to a calibration claim comes from projects that fit temperature scaling afterwards: open-alternative-jev reports calibration error dropping from 5.4% to 2.1% on MMLU that way, and Laya from 0.466 to 0.081 on its English checkpoint. The RLCD guide covers what is actually known about the original method.
A zero-shot classification model takes labels at request time, the same as Jev does, and it has been doing that for years. ModernBERT is the encoder underneath much of this, including Laya's English checkpoint and OpenDecision's NLI head, and it is the baseline Jev's cost and speed get measured against. OpenDecision names its three operations Choice, Noul and Score, which is Jev's vocabulary applied to natural-language criteria rather than to structured question objects.
GLiNER is the odd one out. It finds spans in text rather than deciding between options you supply, so it answers a different question and is only an alternative when what you wanted was extraction. For that job there is now also Clairic, Cloudflare's extraction model, which is a System One model but not a decision model, so it is not in this table. SetFit sits between the two, fitting a classifier from a few dozen labelled examples on CPU. For the detail, see Jev vs BERT, Jev vs GLiNER and the zero-shot classifier definition.
Outlines, Instructor and DSPy already guarantee that an LLM cannot return a type error, which is one of the two things Jev sells. They do it by constraining or validating the output of a model you are already paying for, so the latency and the bill stay the underlying model's. They give you no probability you can threshold on, which is the other half. Structured outputs explains where the line falls, and is Jev just a classifier? asks whether the line is worth a new category.
No. Jev is served only as a hosted API, from TypeSafe AI directly or through OpenRouter. The weights, the parameter count, the architecture and the RLCD training recipe are all unpublished as of 30 September 2026, and there is no announced self-hosting path. What is public is the HTTP API, the documentation and the SDKs, which is what every alternative on this page was built against.
No. There is no weight download and no self-hosted or VPC option, so running a System One model on your own hardware means running something else. The nearest options are the open reproductions: NanoJev and SemIf on a consumer GPU, mini-jev on Apple Silicon or CUDA with about 8.5 GB of memory, Kev on CUDA, ROCm or a 32 GB Mac, CLM on Linux with an NVIDIA GPU, and the encoder-sized models Laya and OpenDecision, which also run on CPU.
That depends on which part you want. For the interface, ekzhang/openjev-sglang implements TypeSafe's POST /v1/systemone endpoints, and Decider and CLM serve the same wire format, so the TypeSafe SDKs work by repointing TYPESAFE_BASE_URL. For a model you can hold and retrain, Laya, Kev, CLM and NanoJev ship their own weights. For a hosted API from another vendor, Together AI serves Tev1, which answers Choice questions only, OpenRouter serves Kev-4B, which answers all three question types, and Upstage, Liquid AI, Inception, Cloudflare and Perplexity serve Solar Decide, d1, Mercury Decide, Clef and pplx-decider, which also answer all three. Laya publishes calibration error figures, and Kev publishes accuracy and Brier scores against Jev on shared eval sets, with its training code. Full comparison: Jev alternatives: hosted, open-source and local options compared.
Not the training method. RLCD has never been described in enough detail to copy, so nothing here reproduces it. Several projects fit temperature scaling afterwards instead: Laya reports expected calibration error falling from 0.466 to 0.081 on its English checkpoint. Others disclaim calibration outright, and the Hugging Face Space named Qwen2.5-1B-RLCD documents no reinforcement learning at all.
No. A check of the OpenRouter models API on 30 September 2026 found no Laya and nothing from Convai Innovations, either among its 464 models or in its separate list of seven decision models, which holds Jev, Kev-4B, Solar Decide and Span-01. Laya is distributed as Apache-2.0 weights on Hugging Face and code on GitHub, and it is meant to be self-hosted. Community ports exist for Node.js and Rust.
38 alternatives
Showing 38 of 38
Declares typed input/output Signatures for LLM modules and optimizes the underlying prompts and weights against a metric. Its Signature abstraction is the closest widely-used open equivalent of Jev's typed-question interface. Star count is GitHub's rounded display figure.
Cua's own family of small System 1 models for computer-use decisions, separate from TypeSafe's Jev. CUA-S1-FORMS is the first release, with weights and a dataset on Hugging Face.
Decision model from Convai Innovations that fully fine-tunes an encoder and adds a typed-decision head: a 421M English checkpoint on ModernBERT-large and a 322M multilingual one on mmBERT-base, with a router picking between them by detected language. Returns choice, score and noul answers, and reports expected calibration error of 0.081 after temperature fitting, down from 0.466 as shipped.
Guarantees valid structured output during generation by constraining decoding to a grammar or schema. The mechanism HN commenters repeatedly cited as already covering Jev's cannot-produce-a-type-error guarantee on ordinary LLMs. Star count is GitHub's rounded display figure.
Pydantic-based library that extracts typed, validated structured data from LLMs across providers with automatic retries. The established way teams get Jev-style typed answers today, at LLM latency and cost. Star count is GitHub's rounded display figure.
A family of small, self-hosted decision models built on Qwen3.5 at 0.8B, 4B and 9B, with training code and an API that matches TypeSafe's System One so TypeSafe's own Python SDK can point at a local server.
Compact encoder model that takes the label set as input at inference time and extracts arbitrary entity types zero-shot on CPU. Same classifier-that-takes-its-categories-as-input shape, for extraction rather than decisions. Star count is GitHub's rounded display figure.
Fine-tunes Sentence Transformer embeddings plus a lightweight head for high-accuracy classification from a handful of labeled examples, with no prompting. The standard cheap alternative to an LLM classification call when the label set is fixed. Star count is GitHub's rounded display figure.
NanoJev is a 0.6B open-source replica of Jev that takes states and questions and returns complete probability distributions with zero output-token decoding. Ships with model weights, dataset and an end-to-end training pipeline, plus a side-by-side demo running Jev, NanoJev and untuned Qwen through Snake and a 50x50 maze.
Modernized BERT encoder used as the backbone for fast fine-tuned classifiers and rerankers. The non-generative baseline Jev's cost and latency claims are usually measured against. Star count is GitHub's rounded display figure.
NeurIPS 2024 masked discrete diffusion language model that produces outputs in parallel rather than token by token. Cited in Jev coverage as the open non-autoregressive research line closest to Jev's parallel sampler.
Python library from Nokia researchers that reads typed Choice, Noul and Score answers with probabilities from an open LLM's logits or hidden states, without fine-tuning. Its zero-label level reduces option-order bias, and a head fit on 100 to 300 labels can run behind vLLM.
Apache 2.0 decision model on a 395M ModernBERT encoder that serves the same /v1/systemone API as Jev, installed as von-sdk from pip or npm. Version 1.2 scores each option on its own, so reordering options no longer changes answers. Its own table puts Von 1.1 at 72.0% on the 49-task jabr v2 suite against 96.6% for Jev.
Deploys DiffusionGemma-Jev (djev) as a Jev API-compatible endpoint on Cloud Run with one RTX PRO 6000 GPU, scaling to zero when idle. The Gemma team's video demo reports 35 to 60 ms single-step latency at roughly $3 an hour.
Google Gemma (@googlegemma)496 starsOpendjev-run: DiffusionGemma-Jev on Google Cloud Run on github.com
Open reproduction of the System One model class, fine-tuned from Qwen3.5-2B-Base, returning a probability distribution per typed question in a single forward pass. Serves TypeSafe's POST /v1/systemone wire format so TypeSafe SDKs work unchanged by repointing TYPESAFE_BASE_URL.
Prefill-only server that implements the TypeSafe Jev HTTP API on open models, running Qwen3.6-35B-A3B on SGLang. Deploys to Modal on a B200 with radix caching and breakable prefill CUDA graphs, scaling to zero after five idle minutes.
A 151M-parameter non-autoregressive decision model, inspired by Jev's approach but trained independently, evaluated on the LocalLLaMA typed-decisions benchmark at 77.10% accuracy, 0.0636 Brier score and 0.0144 expected calibration error.
Educational Jev-like visual inference on Apple Silicon: Qwen3.5-0.8B under MLX answers several questions about one image by reusing shared multimodal context and scoring candidates from logits. Ships a browser UI, CLI, HTTP API and sorting-factory, Breakout and camera-gesture demos.
An open re-creation of Jev built on Qwen3.5-4B that you run on your own GPU. It takes a state and a list of fixed-option questions and answers all of them at once with a probability per option, never free text.
An open-source decision engine inspired by Jev: give it state and typed questions and it scores each option independently, then normalizes the scores into a probability distribution, using any OpenAI-compatible provider key instead of TypeSafe's API.
Independent reimplementation inspired by jevlike that scores pre-written options in one batched forward pass by log-probability over a local Gemma 3 4B, shipping an HTTP server, CLI and a TypeSafe System One API implementation.
A 151M-parameter non-autoregressive decision model on ModernBERT with calibrated uncertainty, inspired by Jev but trained independently, including a benchmark audit against Jev and an in-browser WebGPU playground.
Example server from vLLM PR 57250 that seeds a DiffusionGemma canvas with the answer template and reads one distribution per question slot after a denoise step. It serves a Jev-style question schema over chat completions and is the base djev-run deploys to Cloud Run.
Jev-inspired parallel decisions for CUDA LLMs: independent typed fields are batched through vLLM over one shared context prefix, single-token labels are scored with the pretrained model's own head, and JSON is assembled in code. No retraining. Includes a browser Super Mario run at 71 ms mean inference per action.
Jev-style parallel constrained decisions for any MLX model on Apple Silicon. It takes a schema of booleans, enums and multi-selects, scores every allowed answer for every field in one forward pass, and assembles the JSON itself, so the output is valid by construction and every field carries a probability.
Zero-shot decision engine built on the MoritzLaurer ModernBERT-large NLI checkpoint, around 400M parameters, pulled from Hugging Face on first use. Exposes Choice, Noul and Score over natural-language criteria rather than structured question objects, and states that its confidence figure measures concentration in the returned distribution, not the chance the answer is correct.
Typed, calibrated decisions from any open-weights model in one forward pass on your own GPU, through Hugging Face or vLLM, with no text generated. On RACE-H with Qwen3.6-27B, writing the passage once reached 92.9% accuracy at 4.55 questions per second against 2.00 for batching.
Measures whether an untrained frozen Qwen3-4B can classify against a schema by reading next-token logits for option letters, compared against grammar-constrained JSON generation. Implements analogues of Choice and Noul only; explicitly not a reproduction.
An independent reproduction of Jev's decision layer on Qwen checkpoints, serving the same /v1/systemone request and response schema with choice, score and noul question types, plus an MMLU-Pro benchmark.
Experimental inference harness for typed JSON decisions with DiffusionGemma: denoise freely, then pick the most likely allowed tokens from the final logits, with no training or fine-tuning. Reports results against Jev benchmarks, measured on an H100.
Research preview that takes context, questions and candidate answers and returns probability distributions, running on a frozen Qwen3-4B-Instruct-2507. README states it contains none of Jev's weights or RLCD implementation and claims no performance lead.
Unofficial study reproducing Jev-style parallel constrained decoding on stock 1.5B-8B models on an Apple Silicon laptop. Prefills context once, broadcasts the KV cache across one batch row per schema field and evaluates all fields in a single forward pass.
Open-source, local decision models inspired by Jev and System One, built on Qwen3 and sized for an NVIDIA DGX Spark.
Experimental SwiftUI iPhone app that turns a small vision-language model into an on-device multiple-choice decision tool, reading the next-token logits for A, B, C instead of generating an answer. The displayed percentages are relative scores among the supplied choices, not calibrated probabilities.
An open source library that takes any Hugging Face model and gives it Jev-style structured decision output through an API endpoint, including vision models.
Eugene Cheah - AI builder @ 🇸🇬|🇺🇸 (@picocreator)104k viewsOpenSimple Jev: Turn context into decisions on simplejev.ai
Open-weight decision model built on mmBERT-small that picks one of 2 to 20 supplied options on a CPU, with Apache 2.0 weights on Hugging Face. The launch page scores it against Jev reference values from a published protocol, beating them on three of four tasks and losing on the 72-label Banking77 test.
Supersonic Labs (@supersonicai)OpenJulia 1: a 144M-parameter decision model from Supersonic Labs on supersoniclabs.ia.br39k viewsParameters 144.3MGPU spend ~US$104
HuggingFace Space running a stock Qwen2.5-1.5B-Instruct 4-bit checkpoint that scores multi-field JSON schemas in one pass using KV-cache broadcasting and logit slicing, reporting a 5.6x to 7.0x latency cut on Apple Silicon. The README documents no reinforcement learning or fine-tuning despite the RLCD in the model name.
MIT-licensed PPO model over BAAI/bge-m3 sequence embeddings that emits turn-by-turn conversion probabilities without generating text (arXiv 2503.23303). Its author claims on Hacker News it is the same non-autoregressive, schema-constrained architecture Jev later shipped, a year earlier.
Nothing matches those filters
Clear a filter to widen the search, or send us the thing you were looking for.
