• Updated from E6 to an E8-based model with broader character coverage.
• Increased model size from approximately 3.5M to 5.1M parameters.

• Skip inference for segments without Chinese characters, including recursive child segments.

• Fixed divider misalignment caused by webpage styles.
• Quietly skip Chrome's native error pages.

BZNSYP: all 10,000 sentences; #3 boundaries required, #2 optional, with closing-quote equivalence.

| Functional | Result | Performance | Result |
| --- | ---: | --- | ---: |
| Sentence accuracy | 59.62% | Mean batch time | 34.16 s |
| Precision | 87.71% | Mean throughput | 292.73 sentences/s |
| Recall | 60.07% | Mean backend time per sentence | 24.86 ms |
| F1 | 71.30% | P95 backend time per sentence | 62.96 ms |

Precision counts matched #2/#3 boundaries; recall measures #3 coverage. F1 = 2PR / (P + R).

Measured on Apple M5 Pro, macOS arm64, Node.js v25.9.0: 8 workers, 100 warmup sentences per worker, 2 full passes averaged.

Backend CPU processing only, excluding startup and browser/DOM work. Per-sentence times are measured under the same 8-worker load.

Full changelog

Use the latest trained model for segmentation inference with improved backend rules.

首次支持桌面 Firefox：本地中文分句、开关、分隔线颜色与粗细调节，以及快捷键。