Jev Alternatives: Open-Source Models and Other Options
Last checked · Independent guide, not affiliated with TypeSafe AI
There is no open-weights Jev, but within ten days of launch the community published more than a dozen alternatives. The most popular is Laya, a 421M-parameter Apache-2.0 model that answers Jev-style questions locally. Others fine-tune Qwen models (Kev, Bespoke Nimble, tev1) or wrap an existing open model (SemIf, OpenJev, AnyJev). None is the same model as Jev, and their benchmark claims are their own.
Within ten days of Jev’s launch, developers had published more than a dozen projects that imitate it. They all copy the interface: give the model some text and typed questions, and read back a probability for each option in one pass instead of generating an answer. None of them is Jev. TypeSafe has not released Jev’s weights or training details, so every project below is either a new model trained for the same job or an existing open model wrapped to answer the same way.
Star counts, sizes and licenses below were checked on the projects’ GitHub and Hugging Face pages on September 25, 2026. Accuracy and speed claims are the projects’ own unless we say otherwise; many compare themselves with Jev numbers taken from third parties, not from their own Jev calls.
All 14 projects at a glance
Section titled “All 14 projects at a glance”They fall into three groups. Trained models ship their own weights, trained or fine-tuned specifically to answer Noul, Choice and Score questions. Wrappers train nothing new: they read option probabilities out of an open model you already have, often behind the same POST /v1/systemone interface as TypeSafe. Research projects are experiments and training pipelines rather than ready-to-use models.
The chart shows how far apart they are in size, from Laya and Von at around 0.4 billion parameters to OpenJev at 26 billion. Filter the cards to find ones that fit your hardware.
Showing 14 of 14
- Trained modelApache-2.0
Laya
Nandakishor M, Convai Innovations
- Model
- ModernBERT-large, 421M (multilingual: mmBERT-base, 322M)
- Runs on
- CPU, Apple Silicon or any GPU; pip install laya
The most popular alternative. Ships a server that speaks Jev’s API.
23.7k starsOur head-to-head test → - Trained modelApache-2.0
Kev
Jared Palmer
- Model
- Qwen3.5 / Qwen3.8, 0.8B to 27B
- Runs on
- Apple Silicon Mac up to an 80 GB GPU
TypeSafe’s Python SDK works against a Kev server unchanged; calibrated temperature per checkpoint.
6.8k stars - Trained modelApache-2.0
Von
wfzyx
- Model
- ModernBERT-large, 395M
- Runs on
- GPU or CPU
Version 1.2 makes answers independent of option order.
650 stars - Trained modelApache-2.0
Bespoke-Nimble-9B
Bespoke Labs
- Model
- LoRA adapter on Qwen3.5-9B
- Runs on
- CUDA GPU with BF16
8,192-token context and up to 255 choices per field since Sep 24.
1.8k stars - Trained modelCode MIT; weights not stated
tev1-4B-experimental
Together AI
- Model
- Fine-tuned Qwen3.5-4B
- Runs on
- Your GPU, or Together’s hosted endpoint
Answers with one letter for 2 to 24 options; training recipe published.
110 stars - Trained modelApache-2.0 (per model card)
Jev-Omni
An individual Hugging Face user
- Model
- Gemma 4 12B, fine-tuned on 30,000 questions
- Runs on
- CUDA GPU, about 50 GB in FP32
Takes images, audio and video. Results are self-reported.
209 likes on HF - WrapperMIT
SemIf
TheoLeeCJ (formerly “OpenJev”)
- Model
- Frozen 4B-class open models; 27B bridge
- Runs on
- RTX 3090, Apple Silicon, or the browser via WebGPU
Reads option probabilities straight from an open model.
4.3k stars - WrapperApache-2.0
OpenJev
razorback16
- Model
- DiffusionGemma 26B-A4B
- Runs on
- vLLM on NVIDIA, or MLX on Apple Silicon
Same wire API as Jev; accepts images; also serves Laya. Not the same project as SemIf’s old name.
410 stars - WrapperApache-2.0
AnyJev
Nokia Applied Research
- Model
- Any LLM, e.g. Qwen3-8B
- Runs on
- vLLM
No fine-tuning; fixes calibration and option order with few or no labels.
540 stars - WrapperApache-2.0
simple-jev
Featherless AI
- Model
- Hugging Face open models
- Runs on
- Transformers + PyTorch; free public demo API
Demo API has a 2k-token limit and 2 requests per second.
530 stars - WrapperApache-2.0
Rizzo Flow
Rizzo AI Academy
- Model
- Open models through llama.cpp
- Runs on
- Apple, NVIDIA, AMD or Intel GPU, or CPU only
Same HTTP interface as TypeSafe; probabilities uncalibrated until you calibrate them.
420 stars - ResearchMIT
NanoJev
TianyuCodings
- Model
- Qwen3-0.6B with a full training pipeline
- Runs on
- GPU for training
Demonstrated on ViZDoom, Maze and Snake.
2.2k stars - ResearchMIT
jevlike
vinnylarouge
- Model
- A small scorer you train yourself
- Runs on
- CPU or GPU
Picks among a changing list of options; plays Doom and chess.
1.3k stars - ResearchMIT
mini-jev
r-ms
- Model
- Frozen Qwen3-4B
- Runs on
- GPU for the 4B model
A preregistered experiment reading answer-letter logits.
55 stars
No project matches all of those filters.
Many other repositories use “Jev” in their name but are tools built on Jev, such as the browser agent jev-ultrafast or MCP servers. They need TypeSafe’s API and are not alternatives. Some posts also claim to beat Jev on every benchmark with a small model and no reproducible method; we leave those out until they publish one.
How good are they?
Section titled “How good are they?”No neutral party has tested all of them on the same data yet. What exists:
- Project READMEs. Most compare themselves with Jev using numbers from other people’s Jev runs, with different prompts and sample sizes. Laya’s README says so explicitly.
- JevBench, a community benchmark posted on September 22, 2026, that ranks Jev-class models side by side, including an option-order check. Its diagnostic is why Von 1.2 changed its architecture. It is a volunteer project; see Jev benchmarks explained.
- Our own Jev vs Laya run on 200 public test questions, with both models answering identical questions. Laya won 96 to 86 on a 4-option news task; Jev won 78 to 41 on a 77-option banking task. Details on the Jev vs Laya page.
The pattern across all of them: small open models match or beat Jev on tasks with a handful of well-described options, and fall behind when there are dozens of similar options, long inputs or subtle wording. Most also need calibration on your own data before their probabilities mean what they say.
TypeSafe’s own open-source adapter
Section titled “TypeSafe’s own open-source adapter”TypeSafe publishes system-one-adapter-python, an MIT-licensed package that mimics the official SDK’s system_one call but sends the questions to an LLM from OpenAI or Anthropic. Its stated purpose is comparing TypeSafe with an LLM on cost, speed and quality. It is also a way to keep one codebase while TypeSafe’s signups are paused, as long as you accept LLM prices and latency.
Existing tools that solve similar problems
Section titled “Existing tools that solve similar problems”Jev’s job, choosing among labels from a text description, is an old one. Before Jev, the usual options were:
- Zero-shot classifiers such as GLiClass, which scores text against labels you provide without training. It predates Jev by two years; someone posted it to Hacker News as an open-source Jev.
- An LLM with structured output, asking for one label in JSON. Flexible and easy, but slower, more expensive, and the “confidence” it reports is not a calibrated probability.
- A small fine-tuned classifier trained on your own labeled examples. Fast and cheap at scale, but you need data and a retraining loop every time the labels change.
For a side-by-side of these trade-offs against Jev, see Jev vs classifiers and embeddings and Jev vs LLMs.
How to choose
Section titled “How to choose”| If you need | Consider |
|---|---|
| The real Jev without a TypeSafe account | OpenRouter, Vercel or Cloudflare; see How to get a Jev API key |
| A free, local model that is easy to install | Laya (small enough for a laptop CPU) |
| Local, but closer to LLM-level understanding | Kev-4B or Kev-9B, or Bespoke-Nimble-9B on a GPU |
| Your existing open model, answering Jev-style | SemIf, AnyJev or OpenJev |
| Many languages | Laya’s multilingual checkpoint; Jev itself is best in English |
| Images, audio or video in the question | Jev-Omni or OpenJev, knowing both are young |
| The same code path against an LLM for comparison | TypeSafe’s system-one-adapter |
| Labels scored with no training and no API | A zero-shot classifier like GLiClass |
Several of these (Laya, Kev, OpenJev) accept the same requests as TypeSafe’s API, so you can switch between Jev and a local model by changing the base URL. That makes a head-to-head test on your own data cheap to run.
Will Jev itself become open source?
Section titled “Will Jev itself become open source?”TypeSafe has not announced open weights, a paper or a local version as of September 25, 2026. See Is Jev open source?