Yanlan V3.0 Launches - News

Pre-publication is unforgiving. One slip in a 5,000-word feature, one wrong character in a subtitle that ships to millions, one typo in an official filing — and the cost lands well after the proofreader has gone home. Yanlan V3.0 is built for exactly that moment. It wins all 8 headline metrics over V2.0 and takes 43 of 47 total comparisons. The biggest jump: nearly half the errors V2.0 missed are now caught.

Yanlan V3.0 is the new industry-leading bar for pre-publication Chinese text correction. Missed-error rate falls from 13.06% to 7.04% — a 46% cut — while false alarms also drop, from 1.99% to 1.71%. More real errors caught. Fewer useless interruptions. A proofreader you can actually trust to ship.

Where Yanlan V3.0 fits in

Yanlan V3.0 is built for content operations that can't afford a public mistake: pre-publication review at news organizations, book and journal final-pass at publishers, government communications and policy documents, regulatory and compliance filings at financial and legal institutions, broadcast subtitles and transcripts at media groups. Anywhere a serious piece of Chinese text needs one last rigorous review before it goes out, V3.0 belongs in the loop.

Tuned for the hardest errors, not the easy ones

V3.0 doesn't optimize for casual typos. It's tuned for the cases that actually trip teams up at the finish line — subtle wording errors, context-dependent corrections, domain vocabulary, and corrections where the model has almost no room to be wrong. We rebuilt the high-difficulty Chinese correction benchmark from scratch and upgraded the data synthesis engine, so V3.0 trains against text that looks much closer to what real editorial and compliance desks see.

Catches more. Missed-error rate down 6.02 points. C-Recall up 6.02 points. Even subtle, context-dependent errors are surfaced.

Fewer false alarms. False-alarm rate continues to fall. C-Precision holds at 98.30%. Review teams stay focused on real issues.

Lands the right fix. Perfect correction rate jumps from 84.66% to 94.16% — a 9.5-point leap. Detection is the straightforward half; producing a correction an editor can accept is where the real value sits.

Eight metrics. Eight wins.

Weighted across three high-difficulty sample sets (484 / 991 / 2933), V3.0 sweeps every headline number: Accuracy, F1, F0.5, missed-error rate, false-alarm rate, correction precision, correction recall, and perfect correction rate. No cherry-picking, no quietly avoided weak spots.

Three high-difficulty sample sets (n=484 / 991 / 2933) weighted by sample size. V3.0 leads on all 8 metrics with zero regressions.

95.74%Accuracy · V3.0 (+3.32 pp)

7.04%FPR · V3.0 (−6.02 pp)

1.71%FNR · V3.0 (−0.28 pp)

94.16%Perfect correction · V3.0 (+9.50 pp)

43 out of 47. System-level lead.

Pre-publication correction is never a single-metric contest. Detection, correction, false-alarm control, production stability — any one of them slipping is an on-publish risk. V3.0 wins 43 of 47 comparisons across every dataset and every metric we tracked. That's system-level capability, not a single point lucking into the spotlight.

Broken out: V3.0 leads 35–4 across the 39 single-dataset × single-metric pairs and sweeps the 8 weighted-aggregate metrics 8–0.

The biggest leap is in the fix itself.

Spotting that something looks off is comparatively cheap. Producing a fix your editor will actually accept — that's where most proofreaders fall apart. It's also where V3.0's biggest leap lives: perfect correction rate clears 93% on every dataset evaluated.

Yanlan V3.0 Yanlan V2.0

Across three high-difficulty evaluation tasks (Task 1: n=484, Task 2: n=991, Task 3: n=2933), perfect correction rate moves up in lockstep — the +9 pp floor holds on every task, not propped up by a single one.

The full evaluation was conducted on three internal, high-difficulty tasks (n=484 / 991 / 2933) representing a range of pre-publication content types. Per-task detail is available on request from the Coolwei team.

Yanlan V3.0 is now fully available

A new industry-leading bar for pre-publication Chinese text correction. Eight headline metrics, eight wins. 43 of 47 comparisons won. Missed-error rate down to 7.04%, perfect correction rate at 94.16%. Already serving news and publishing organizations, government communications, financial and legal compliance, and broadcast media. Enterprise and government teams: contact us to evaluate Yanlan V3.0 on your own content.