A 16 GB GPU
Winnow-12B was tested in its 8-bit build on a 16 GB card, and the same server also chats and reads images.
Jev local & open-source guide · updated October 4, 2026
Can you run Jev locally?
No. TypeSafe’s Jev is a proprietary, hosted model. Its weights are not on Hugging Face, GitHub or Ollama, so it cannot be self-hosted. What you can run locally are open-weight models that answer the same typed questions. This page lists them with their licences, hardware needs and JevBench v1.6.0 standings.
Is Jev open source?
The model is closed. Some of the tooling around it is public.
- Model weights: not released. TypeSafe has not disclosed Jev’s size, base model or training data.
- Access: a hosted API. Your text is sent to the service that runs the model.
- SDKs: TypeSafe publishes client libraries such as
@typesafe-ai/sdk. They call the hosted API; they do not contain the model. - “OpenJev” and similar names: independent projects on open base models. At least ten unrelated projects use that name. See which OpenJev is which.
Facts about Jev are from TypeSafe’s public documentation as of September 18, 2026.
Open models you can run locally instead
Each answers Jev-style yes/no, choice and score questions. Standings are from JevBench v1.6.0, where hosted Jev is #2 at 71.1.
| Model | Licence | Hardware and speed notes | JevBench v1.6.0 |
|---|---|---|---|
| WinnowEldanRing | Apache-2.0, with the Gemma 4 terms that apply to derivatives | A 16 GB GPU for the Q8 build; the BF16 build needs more memory or offloading | #6 of 92 · 68.9 |
| Cygnetblockbrain | MIT shim; Gemma 4 weights under Apache-2.0 and Google’s prohibited-use policy | A 48 GB GPU in the author’s runs; other cards not measured | #7 of 92 · 68.6 |
| Jev-Omniakhilaaa3 | Apache-2.0 weights on Gemma 4 12B | A CUDA GPU; about 24 GB of BF16 weights | #8 of 92 · 67.7 |
| decider-4bMapika | Apache-2.0 for weights and package | One CUDA GPU, about 8.4 GB in bf16; also runs on CPU | #17 of 92 · 41.2 |
| JevK5allebee | Apache-2.0 for weights and code | Your own GPU, about 9GB in bf16; the author’s figures are on an H100 and an L40S | #18 of 92 · 37.4 |
| Imajev-4BMohit Garg | Apache-2.0 adapters and code; Apache-2.0 Qwen base | Apple silicon through MLX or one CUDA GPU; 9.3 GB base weights for the 4B | #26 of 92 · 28.7 |
| SemIfTheoLeeCJ | MIT for the project code; model weights keep upstream licences | Your own GPU; the published numbers are on an RTX 3090 | #45 of 92 · 15.5 |
| LayaConvai Innovations | Apache-2.0, weights on Hugging Face, installable from PyPI | 32.8–39.5ms on a Tesla T4; 193–464ms on CPU | #84 of 92 · 0.0 |
| djevMaisa | Community djev-dev repo is Apache-2.0; Maisa’s hosted service is not yet released | Possible: roughly one NVIDIA B200, Linux, CUDA 13, BF16 | Not measured |
| OpenJevrazorback16 / Codiv | Apache-2.0, weights from NVIDIA and Google, about 18GB | 24GB NVIDIA GPU under vLLM, or Apple silicon with 16GB free under MLX | Only in thinking mode, #40 |
| Mercury DecideInception | See the comparison | See the comparison | Not measured |
Licence and hardware details are quoted from each project’s own repository or model card, as read on October 6, 2026; follow a model’s link for the sources. The JevBench composite mixes intelligence, calibration, speed and cost, so a fast small model can rank near Jev while answering fewer decisions correctly. Read how JevBench scores are built.
Pick by the hardware you have
Start from your machine, then check the comparison for that model.
About 8–9 GB of GPU memory
decider-4b and JevK5 are Qwen3.5-4B models of that size. SemIf publishes its numbers on an RTX 3090.
Apple silicon
Imajev-4B runs through MLX, and razorback16’s OpenJev runs under MLX with 16 GB free.
A large GPU
Cygnet serves an unmodified Gemma 4 12B on vLLM and was measured on 48 GB and 96 GB cards. Jev-Omni also needs a CUDA GPU with room for a 12B model. JEMM, an Apache-2.0 Qwen3.8-27B decision model that also reads up to four screenshots, asks for a 64 GB card and has no JevBench score yet.
CPU only
Laya is a 421M encoder that runs on a CPU, but it needs fine-tuning on your labelled data before it is useful.
No hardware
Use hosted Jev. Jev AI also hosts Laya, Clef and Jev-Omni in the same playground, so you can compare before you install anything.
Hosted Jev or a local model?
The choice is mostly about data, volume and how much accuracy you need.
Reasons to run locally
- The data cannot leave your network.
- You already own GPU capacity, or volume is high and steady enough to keep one busy.
- You need millisecond latency on your own hardware, offline use, or weights you can inspect and fine-tune.
Reasons to stay hosted
- Nothing to install, operate or keep warm.
- Jev scores 62.4 on intelligence in JevBench v1.6.0; the open 4B models score well below it.
- A 64k-token context, where several local options stop at 16k or 32k tokens.
- Probabilities that are calibrated as delivered. Several open models document a calibration step you must fit on your own data.
Moving existing Jev code to a local model
The request shape often carries over. The behaviour does not.
Several alternatives accept the TypeSafe-style /v1/systemone request: decider-4b and OpenJev are built to work with TypeSafe’s SDKs after a base-URL change, and JevK5 and Cygnet expose the same route. Jev AI’s own endpoint works the same way, which is how the SDK setup points an existing client here.
Before you switch, check each model’s option and context limits, rerun your labelled examples and re-fit any confidence thresholds. A threshold tuned on Jev does not transfer to another model.
Running Jev locally: FAQ
Can you run Jev locally?
No. Jev is a proprietary model that TypeSafe serves through a hosted API. TypeSafe has not published the weights, so there is nothing to download and run on your own machine. You can run an open-weight alternative locally, or call hosted Jev.
Is Jev open source?
No. The Jev model is closed. TypeSafe publishes client SDKs, but open client code does not make the remote model open source. Projects with names such as OpenJev are independent rebuilds on other base models, not a release of Jev.
Can I download Jev from Hugging Face or GitHub?
Not the TypeSafe model. Repositories that carry the Jev name on Hugging Face or GitHub are community models, such as Jev-Omni or JevK5, built on open bases like Gemma 4 or Qwen. Check the author and the base model before assuming it is Jev.
Does Jev work with Ollama?
There is no official Jev build for Ollama, because the weights are not public. Some open alternatives ship in formats local runtimes can load; Winnow-12B, for example, is released as GGUF with its own llama.cpp-based server.
What is the best open-source alternative to Jev?
It depends on your hardware and inputs. In JevBench v1.6.0, Cygnet and Winnow-12B place above Jev overall, while the 4B models are faster and lighter but trail Jev clearly on intelligence. Test the candidates on your own cases before choosing.
How large is Jev, and what hardware does it need?
TypeSafe has not disclosed Jev’s size or base model. Because it is an API, you need no GPU to use it. The open alternatives range from a 421M encoder that runs on a CPU to models of 12B and more that need a dedicated GPU.
Will my Jev code work against a local model?
Often, with a changed base URL. Several alternatives accept the same /v1/systemone request shape. Interface compatibility is not model equivalence, though: limits, calibration and answer quality differ, so re-validate your thresholds.
Set a baseline with hosted Jev first
Run your own cases in the playground, keep the answers, then compare a local model against them.
Every model named here belongs to its respective author and is not affiliated with Jev AI. Jev AI is independently operated and is not TypeSafe’s official site. Figures can change; the linked comparison pages list their sources.