These past two days, the popularity of the Jev model has

Classified 500 People's Daily articles for Hubei-related content with Jev, cutting per-article time from 3 seconds to 0.35 seconds versus Gemini Flash Lite.

The original post, translated

These past two days, the popularity of the Jev model has been extremely high, and it can even be said to be the highest density of original content I've seen since I started focusing on English Twitter in the past half year.

Coincidentally, over the past year, I've been running a small project, and one part of it is to classify the news published daily in the People's Daily, determining whether each article contains Hubei-related elements.

After running for a year, I've accumulated nearly 24,000 data entries, of which about 1,800 were judged to contain them. Based on this data, I randomly selected a test dataset, which includes 500 entries judged to contain and 500 entries judged not to contain.

Previously, the model I used was always the Gemini Flash Lite series, which after testing was considered the best choice considering both cost and speed.

Today, after OpenRouter launched this Jev, I used the same prompt to make a small demo for a comparative test. The video below is a live recording of running 16 processes concurrently.

I have to say, the test results exceeded expectations and were very pleasantly surprising.

First, the speed is really fast! Compared to Gemini Flash Lite's average time of 3 seconds per article, Jev reduced the time per article to 0.35 seconds. This nearly 10-fold speed increase truly achieves a qualitative leap, bringing a lot of imaginative space for new scenarios.

Second, the quality is also really good! Compared with Gemini Flash Lite's classification results, Jev showed inconsistent judgments on about 15% of the content.

But after my manual review, most of the differences were due to personal names. For example, when the name of an official who previously served in Hubei Province appears in the news, Flash Lite, due to its larger knowledge base, would judge it as containing, while Jev is weaker in this area and thus judged it as not containing. In my actual application scenario, this kind of judgment difference is completely acceptable.

Finally, the price is also really cheap! Looking at the OpenRouter bill, I ran about 5,000 data entries back and forth, 30M tokens, and the total cost was only 1 dollar! With this cost-effectiveness, what more could you ask for!

In short, I personally think that the classifier large model represented by Jev is indeed a very promising development direction, and it can even be said to have achieved the impossible triangle of speed/quality/price in classification scenarios. Friends with similar needs can really consider it carefully.

That's all for sharing. I have to stay up all night using it to optimize my production project.

The post above is a machine translation from zh; the untranslated text is in the fold-out.

Engagement when collected

Numbers are a snapshot taken from X when the case was added to the library (schema v1, collected 2026-09-19); they will not match today.