What is new in TeaNovel summer 2026 — improved NER for named entity consistency, expanded platform support, and what we are building next for 2026.
The TeaNovel Summer 2026 update brings three concrete features: a rebuilt named-entity recognition layer that keeps character names consistent across chapters, browser-extension support for two additional novel platforms, and a faster SSR rendering path for the in-browser reader. Here is what changed, why we built it, and what we are shipping next.
The single most common complaint in our feedback queue has been character-name drift — the protagonist is "Xiao Yan" in chapter 1, "Little Flame" in chapter 40, and "Senior Brother Yan" in chapter 112. All three are technically correct renderings of different address forms in Chinese, but without a consistent anchor they read as three different people.
We rebuilt the NER pipeline from scratch this spring. The new system does two things the old one did not:
Entity anchoring. The first time a name appears in a novel, we log the canonical romanization and the base Chinese characters. Every subsequent chapter runs a reconciliation pass before the translation is finalized, matching address variants back to that anchor. "小炎", "炎哥", and "炎兄" all resolve to the same entity record.
Reader-visible glossary. Your novel terminology management settings now show the auto-detected entity list for each novel. You can override any canonical name — if you prefer "Yan Xiao" over "Xiao Yan", change it once and the engine reprocesses cached chapters automatically.
This matters most for cultivation novels with large ensemble casts. We tested the new system against a 300-chapter xianxia title and saw name-inconsistency errors drop by around 70% in our internal testing compared to the previous build. We will keep measuring — the metric is part of our internal quality dashboard now.
The browser extension now supports Zongheng (纵横中文网) and SF Acg (SF轻小说), joining the existing adapters for JJWXC, Qidian, Fanqie, and Ciweimao.
Zongheng: The adapter handles both the standard chapter view and Zongheng's VIP chapter paywall detection. If you have a paid Zongheng account, the extension reads the unlocked chapter content. If not, it surfaces a clear "chapter locked" notice rather than feeding partial text to the translator. Partial-text translations are worse than no translation — we would rather be honest about the limitation. See how to read Zongheng novels in English for the full setup guide.
SF Acg: SF has a different HTML structure from the major mainland platforms and uses a chapter-page model with aggressive rate limiting. Our adapter throttles requests automatically. Translation of very long SF chapters (some run 6,000+ Chinese characters) may take 15–25 seconds; we surface a progress bar rather than a spinner so you can see it is working. See how to read SF Acg novels in English for the setup guide.
If you have noticed the in-browser reader loading faster on first open, that is the new SSR caching layer. Previously, server-rendered reader pages recomputed layout on every cold request. We now cache the rendered output per chapter per user configuration (font size, night mode, column width) and serve it directly. Median first-contentful-paint for chapter pages dropped from around 420ms to 140ms in our internal benchmarks on a warm CDN node.
This does not change how translation works — it only affects how fast the already-translated chapter page appears. For the translation pipeline itself, the flow is still: cache check → credit freeze → streaming SSE → write cache → credit deduction. Nothing in that chain changed.
As of the summer 2026 update, TeaNovel's library has 134 novels across danmei, xianxia, xuanhuan, romance, and sci-fi categories.
Translation pricing is unchanged: 25–35 credits per chapter, depending on chapter length. Longer chapters (typical of Qidian serialization, which can run 4,000+ characters) land at the top of that range. Short chapters on JJWXC or Fanqie usually land at 25.
New accounts get 1,000 credits each month — enough for 25–40 average-length chapters. We know that is a generous taste. The intention is to let you verify that the translation quality works for your specific novel before you commit. If the first chapter of your target title comes through cleanly, the rest of the novel will too; the engine has behaved deterministically in our testing on a per-novel basis once the entity glossary is built.
Being transparent about the roadmap means being transparent about what is not done yet.
We are still not calling Fanqie production-ready. The adapter works, but Fanqie's anti-scraping layer is aggressive and we see a higher-than-acceptable timeout rate on peak traffic hours (roughly 8–11 PM CST, based on our internal operational observation). We are not promoting it as a first-class platform until that stability issue is resolved. If you use Fanqie and hit timeouts, the retry button in the extension sends a fresh request — it is not a dead end, just slower. See how to read Fanqie novels in English for the current workarounds.
Bulk translation for full novels is in private beta. A handful of users are testing the ability to queue an entire novel for batch translation overnight. We expect to open that to all users later this summer. In the meantime, the existing bulk translation guide covers the manual multi-chapter approach.
Mobile reader improvements are next on the interface roadmap. The desktop reader got significant work this cycle; mobile did not. Our target is to close the gap on tap-target sizing and column layout before end of Q3 2026. We will share concrete numbers once the redesign is testable.
We get asked regularly how TeaNovel compares to running a novel through GPT or Claude directly. The honest answer: the raw model quality is similar, but the context management is not. A standalone API call has no memory of chapter 1 when it is translating chapter 50. Our pipeline maintains a per-novel context window that includes the entity glossary, a style profile built from the first ten chapters, and inter-chapter continuity signals. That is the actual product — not the model, but the scaffolding around it.
For a deeper look at how the engine itself works, the NoveLM engine explainer goes into the architecture.
Three shipped this cycle: rebuilt named-entity recognition for consistent character names, browser-extension adapters for Zongheng and SF Acg, and an SSR cache layer that cuts reader first-contentful-paint from around 420ms to 140ms. The roadmap sections above cover what is still in progress.
Yes, with a caveat. When you open a novel that has existing cached chapters, the system detects missing entity records and runs a background reconciliation job. Chapters that were already cached are not automatically re-translated — the old text stays — but new chapters from that point forward use the updated glossary. If you want the full novel re-translated with the new NER, you can clear the chapter cache from the novel settings page and retranslate. That will consume credits.
The browser extension has stable adapters for JJWXC, Qidian, Ciweimao, and the two new additions this cycle, Zongheng and SF Acg. Fanqie is supported but flagged experimental due to stability issues. Haitang (海棠) is on the roadmap but does not have a working adapter yet — the platform's structure makes it significantly harder than the others.
Go to the novel's detail page, open the Terminology tab, and find the entity in the auto-glossary. Click the name to edit the canonical romanization. Changes apply immediately to new translations. For existing cached chapters, you will need to clear and retranslate if consistency across the full archive matters to you.
The blog is the primary channel for product announcements. We also post release notes in the in-app notification center when changes affect pricing, credits, or extension behavior. For community discussion, the Novel Updates forum and related communities are where most readers compare tools — we read those threads, even when the feedback is critical.
That is the summer 2026 update. The NER work in particular took longer than we planned — named-entity consistency in Chinese fiction is genuinely hard because the same character can be addressed by a title, a given name, a courtesy name, and a pet name across a single novel, all carrying different relationship signals. We think the new system handles this substantially better. Tell us when it does not.
Start Reading with 1,000 Free Credits
No credit card. 1,000 credits refresh every month — roughly 25–40 chapters. Install the extension and translate any Chinese novel.