Hassan trained tev1-4B, a Jev-like classifier, for $17
Weights, data recipe and a training tutorial are out, with no accuracy comparison against Jev posted yet.
Hassan announced tev1-4B-experimental, a Jev-like classifier fine-tuned on top of Qwen3.5 4B, and said the training run cost $17. He published the weights, the data recipe and a full tutorial for fine-tuning your own small classification model, plus the code. His suggestion for anyone who does not want to read the tutorial is to point an agent at the repo and let it run the fine-tune.
I'm releasing everything: the weights, data recipe, & a full tutorial on how to train your own.
The hosted version is live on Together serverless at $0.042 per 1M input tokens and $0 per M output tokens. That is the price of the model as served, not a measured cost per task, and nobody has posted a workload where it was put up against Jev on the same inputs.
The schema question
ZazenCodes asked the two questions that matter for anything meant to sit in a Jev-shaped slot, inference times and whether it carries the same schema guarantees, specifically no hallucinated classification labels. Hassan's answer was that when you use the model with regex, which he calls the recommended way, it should always adhere to the schema. That is constrained decoding doing the work, not the model being trained out of inventing labels, so the guarantee is a property of how you call it rather than of the weights.
On latency he did not give a number. He said end to end latency will depend on how many people are hitting the model at the same time, and invited ZazenCodes to try it. For a shared serverless endpoint that is an honest answer and also a warning, the figure you measure today is not necessarily the figure you get in production.
What is missing
There is no accuracy comparison against Jev in the announcement. There is no task suite, no confusion matrix, no per-label breakdown, and no latency distribution. What exists is a cheap training recipe, an open artifact and a serving price, which is enough to reproduce but not enough to substitute.
For people building on it, the interesting number here is $17, not the accuracy. The claim being made is that a classifier of this shape is cheap enough to train on your own labels rather than rent, and the release is the evidence for that claim. Whether tev1-4B itself holds up against Jev on your task is still an open experiment, and the first person to run it on a real workload will have more to say than the launch post does.
when you use it with regex (which is the recommended way), it should always adhere to it