Jev is built for decisions that happen constantly inside AI applications: route a request, choose an agent action, score a condition, or flag a policy issue. It takes application state plus a predefined question and returns a typed answer with probabilities that code can use immediately.
For developers who want that workflow on local hardware or inside a self-hosted stack, there are now several open-source options worth a look. Here are ten of the best Jev alternatives available in 2026.
Last updated: October 26, 2026.
Quick comparison
| Project | Best for | Deployment |
|---|---|---|
| Clef / Clef-flash | Jev-compatible multimodal decisions with local or hosted deployment | NVIDIA GPU or Workers AI |
| GLiNER2.5-Decide | 340M CPU/GPU classification with cross-question constraints | CPU or GPU via gliner2 |
| Laya | Multilingual non-autoregressive decision models | CPU, GPU, or Apple MLX/Core ML ports |
| Kev | Trainable Qwen models with a System One API | CUDA/ROCm, CPU, or Apple Silicon |
| d3 / d3-turbo | Text, image, and video decisions with multi-question input sharing | GPU or hosted Decision 3.0 API |
| SemIf | Direct option scoring from open language models | NVIDIA GPU, Apple Silicon, or browser WebGPU |
| mammoth-1 / elephant-1 | 27B and 9B multimodal decisions with hosted or local deployment | CUDA GPU or Baseten API |
| Bespoke Nimble | Qwen-based typed decisions with a published training recipe | NVIDIA GPU or Apple Silicon |
| Mapika Decider | Jev-style decisions with GGUF, CPU, and GPU deployment | CPU, CUDA, Apple MPS, or GGUF/Metal |
| jevos | Jev-compatible typed decisions on a laptop CPU | CPU native binary or GGUF |
Best Open-Source Jev Alternatives
1. Clef / Clef-flash
Clef and Clef-flash are Cloudflare’s open decision models. Clef is a 27B model built on Qwen3.8-27B, and Clef-flash is a 9B model built on Qwen3.5-9B. Both accept text, JSON, images, or video as state and return probabilities for choice, score, and noul questions in one forward pass.
Their request and response format is compatible with Jev and System One. Cloudflare hosts both models on Workers AI, and the Apache-2.0 weights are available on Hugging Face for local deployment.
Clef uses a joint schema head to score all allowed options across the requested questions. Clef-flash keeps the multimodal input and Jev-compatible API in a smaller 9B package for lower-latency deployments. Self-hosting requires memory for the model backbone and vision encoder, in addition to the small decision head.
License: Apache-2.0
2. GLiNER2.5-Decide
GLiNER2.5-Decide is Fastino’s 340M decision model built on a DeBERTa-v3-large encoder. You pass text and a label schema at call time, and the model scores the labels in a single forward pass with no generated text.
It runs through gliner2 on CPU or GPU and can score several decision heads against one input. Cross-question constraints can enforce rules between related outputs, including cases where one answer limits another.
Fastino’s 17-domain evaluation reports 60.2% average exact-match accuracy for GLiNER2.5-Decide, compared with 56.4% for SemIf and 46.6% for Laya under that test setup. The released checkpoint is designed for English text. Its gliner2 label-schema calls require adaptation from a Jev client.
License: Apache-2.0
3. Laya
Laya is a non-autoregressive System One decision engine for choice, score, and noul questions. One forward pass handles the application state and requested questions, returning typed answers with probabilities.
Laya has a 421M English checkpoint, a 322M multilingual checkpoint for 100+ languages, and a 421M typed-decisions checkpoint. Its Router selects the appropriate checkpoint from the input language. Default context budgets are 512 tokens for English and 1,024 for the multilingual and typed-decisions checkpoints. The question and options share that budget with the state.
Laya can be fine-tuned on custom data. Its training package includes RLCD training, temperature fitting, evaluation, and Hugging Face publishing. The typed-decisions checkpoint reached 0.766 accuracy on Laya’s 2,000-decision evaluation. Laya-MLX runs these checkpoints natively on Apple Silicon, and a Core ML port is also available.
License: Apache-2.0
4. Kev
Kev is a family of Jev-style decision models built on Qwen3.5 and Qwen3.8. The local server matches TypeSafe’s System One request format, and TypeSafe Python SDK clients can connect after changing the base URL.
The current family has 0.8B, 4B, 9B, and 27B checkpoints. Kev-4B is the recommended starting checkpoint. Kev-27B uses Qwen3.8-27B and is intended for an 80 GB GPU. Kev-0.8B and Kev-4B have documented Apple Silicon configurations. The 0.8B, 4B, and 9B checkpoints also run on CUDA GPUs.
Kev includes the data pipeline, evaluation suites, training code, and local server for custom decision models. Kev-27B reached 0.889 test accuracy on Kev’s current New Sources test set. For deadline rules, include elapsed day counts in the state because date subtraction is unreliable.
License: Apache-2.0
5. d3 / d3-turbo
d3 and d3-turbo are 27B multimodal decision models from vLLM Semantic Router. They accept text or JSON with images and videos, then return probabilities for choices, yes/no answers, and scores without generating text.
d3-turbo reads the input once per request and answers multiple questions about it. It also returns bounding boxes for images and video frames, or tracks an object through a video. Its default context is 32,768 tokens per question, with up to 131,072 tokens when long_context=True.
Both models have downloadable weights and inference code for self-hosting. d3-turbo can also be tried through the hosted Decision 3.0 API. Local deployment requires a GPU setup sized for the 27B model. BF16 weights alone occupy about 52 GB before runtime and input-processing memory.
License: Apache-2.0
6. SemIf
SemIf uses existing open language models for typed decisions. It reads the scores assigned to declared answer options directly from model logits and returns typed probabilities with no answer sentence or JSON generation step.
SemIf supports Qwen3.5-4B and MiniCPM5-2B in its local and browser examples. Recent updates added Apple Silicon scoring through PyTorch/MPS, a Qwen3.8-27B EXL3 bridge, and per-workload temperature calibration.
Its WebGPU demo runs quantized models directly in the browser. A shared application state can also be prefetched and reused across multiple questions. Memory requirements and decision quality depend on the base model and quantization you choose.
License: MIT
7. mammoth-1 / elephant-1
mammoth-1 and elephant-1 are decision models from Base Labs at Baseten. The 27B mammoth-1 is built on Qwen3.5-27B, and the smaller 9B elephant-1 uses Qwen3.5-9B. Both accept text and images.
Each request can contain several choice, noul, and score questions. The models process the shared state once and return a probability distribution for each question. This supports tasks such as ticket routing, refund checks, and urgency scoring in one request.
Both models support sequences up to 16,384 tokens and reject requests that exceed their length limits. They are available through Baseten Model APIs or as downloadable weights with a CUDA-based self-hosting runtime.
License: MIT
8. Bespoke Nimble
Bespoke Nimble is a Qwen3.5-9B decision model and training recipe for choices, booleans, and rubric score levels. It scores allowed answer tokens directly and returns a probability distribution. It does not generate free-form text.
The latest Bespoke-Nimble-9B release supports an 8,192-token context and up to 255 choices per field. The project publishes its training examples, held-out examples, inference code, and evaluation material. It runs on Apple Silicon or an NVIDIA GPU. The Hugging Face release is a LoRA adapter and requires the Qwen3.5-9B base weights as well.
The Hugging Face model package is Apache-2.0. The GitHub repository does not currently include a root LICENSE file.
License: Apache-2.0 for the released model package
9. Mapika Decider
Decider is Mapika’s family of decision models built on Qwen and Gemma backbones. It scores predefined answers in one forward pass and returns probabilities for choice, score, and noul questions. Its HTTP server exposes the Jev-style /v1/systemone endpoint.
The family includes 2B and 4B models for smaller machines, plus larger 12B and 31B options. The 2B and 4B GGUF releases run through llama.cpp on CPU, CUDA, or Metal. Their Q4_K_M files are about 1.3 GB and 2.7 GB, respectively. These GGUF builds are text-only and need Decider’s option-scoring runtime. Loading one as a regular chat model does not reproduce its decision API.
Decider also includes training code, datasets, temperature calibration, and evaluation tools for custom decision models. The PyTorch runtime can use CUDA, Apple MPS, or CPU, and image-capable checkpoints are available alongside the text models.
License: Apache-2.0
10. jevos
jevos is a CPU-first local decision server built around a compact decision model. The current jevos-v5 release runs 8-bit weights through OpenVINO in one native binary with no Python or GPU required.
The server speaks TypeSafe Jev’s wire format and supports noul, choice, and score requests. Choice and score questions are resolved through the model’s yes/no decisions. GGUF releases are also available for llama.cpp. jevos-v5 supports an 8,192-token context, and score support is early.
The earlier jevos-v3 checkpoint answered 80.8% of 999 hand-written yes/no questions correctly and 86.5% of 821 new questions written by another model family.
License: MIT
Quick picks by deployment need
- CPU with no Python setup: jevos-v5 runs as a native binary on a laptop CPU. Start with yes/no decisions before evaluating its early score support.
- Apple Silicon: Laya-MLX with the 421M English Laya checkpoint offers native MLX inference for short text decisions. Use the multilingual checkpoint for other languages.
- Existing Jev Python clients: Kev-4B accepts the System One request format and works with the TypeSafe SDK after changing the endpoint and model settings.
- Images and video: d3-turbo answers multiple questions about shared visual input and can return boxes or tracks. Its hosted API avoids a local 27B deployment.
Switching from Jev to a local model
Clef, Kev, and Decider expose the Jev-style /v1/systemone request format. jevos also accepts it, with unchanged SDK use for yes/no questions. GLiNER2.5-Decide uses gliner2 calls, and Nimble uses its own scoring helper.
Matching the API does not guarantee matching answers or confidence values. Replay a held-out sample of your requests before switching, then fit decision thresholds to the new model. Include long states, ambiguous options, and inputs with missing evidence in that sample.
With Kev-4B running at http://127.0.0.1:8009 and the TypeSafe SDK installed, this client sends a local yes/no request. Keep the state and question definitions you already use, and change the endpoint, API key, and model settings for your server:
from typesafe_sdk import Noul, TypeSafeClient
client = TypeSafeClient(
api_key="local",
base_url="http://127.0.0.1:8009",
model="kev-latest",
)
result = client.system_one(
state="Refund the duplicate charge.",
questions={
"billing": Noul(
instructions="Is this a billing issue?"
)
},
)
print(result.nouls["billing"].noul)FAQs
Is Jev open source?
No. Jev is a hosted TypeSafe model, and TypeSafe has not published its model weights. The alternatives in this list provide open weights, source code, or both for local Jev-style decisions.
Which open-source alternative is closest to Jev?
Clef and Clef-flash document full Jev and System One API compatibility. Kev is designed as a drop-in System One replacement. jevos also speaks the Jev wire format and supports noul, choice, and score requests on a local CPU. These are API similarities, not a guarantee of equivalent accuracy or probability calibration.
Which Jev alternatives work on Apple Silicon?
Laya has Apple-focused ports, Kev supports Apple Silicon for its smaller checkpoints, SemIf includes PyTorch/MPS support, and Bespoke Nimble can run on Apple Silicon. Decider also runs through PyTorch/MPS or through llama.cpp with Metal.
Which Jev alternatives can run on a CPU?
GLiNER2.5-Decide and jevos are built for CPU inference. Laya also supports CPU inference, and Decider’s smaller models can run on CPU through PyTorch or llama.cpp.
Can these Jev alternatives run fully offline?
Yes. Every entry in this list has a local deployment path after the required model files and runtime are downloaded. Clef and Clef-flash are also available as hosted models on Cloudflare Workers AI.
Related Jev resources
- The Ultimate Jev Resource List: SDKs, Agents, MCP Servers & More
- 10 Jev Use Cases: Real Projects & Demos
Changelog
October 26, 2026: Replaced Laya-MLX, NanoJev, and jevlike with d3 / d3-turbo, mammoth-1 / elephant-1, and Mapika Decider.
