Yanlan model upgrade: full-modality Chinese correction goes live

Three hard numbers, moved together

The upgrade resets the three figures that decide what deployment costs. The false-alarm rate falls from 5% to 0.5%. Single-machine throughput rises from 2,000 to 30,000 characters per second. Full-system deployment cost falls from the H100 tier at ¥2,000,000 to the RTX 4090 tier at ¥200,000, a 90% reduction.

6-stage mixed training and routed processing

Yanlan is trained in 6 mixed stages of supervised fine-tuning (SFT) and reinforcement learning (RL), tightening the model's detection and correction boundaries stage by stage. At inference a classifier predicts the text type — government, legal, media — and routes the input to the central dictionary, the correction model, or a custom dictionary, before post-processing and the result integration layer produce the output. That routing takes overall complexity from O(n²) to O(nlogn + kn), raises parallel capacity about 30×, and cuts memory use by 60%.

The pipeline is layered by precision: L1 lexical analysis at <1ms per thousand characters, L2 grammar detection at 30ms, L3 semantic understanding at 500ms. Each layer is optimised on its own, coarse filtering and fine correction stay separate, and the false-alarm rate holds at 0.5%.

Every modality in one system

Beyond text, video subtitles, audio transcripts and face checks run in the same system, sharing one dictionary and one post-processing chain.

Billion-scale training data and authoritative dictionaries

Training data moved from millions to billions of samples, a 1000× increase, drawn from authoritative media and official documents and refreshed daily, so error patterns are covered far more completely.

On the regulatory side the full national laws and regulations database is loaded — 16,740 documents covering the constitution, statutes, administrative regulations, local regulations and judicial interpretations — synchronised daily with the official source, with repealed and amended provisions detected automatically so historical text is never cited. The officials dictionary is collected daily from the Central Commission for Discipline Inspection site, new entries land within 24 hours, and images, video and documents can be checked in batch with position changes updated automatically.

On-premise deployment and production reliability

All data is processed locally and never leaves the internal network, cutting end-to-end latency 90% against a cloud round trip. Hardware runs on consumer cards such as the RTX 4090 and 5090 and on domestic silicon including Huawei Ascend and Hygon DCU; a single machine within a ¥200,000 budget carries tens of millions of characters per hour. Three deployment shapes are available: on-premise (data stays inside the network, highest security tier), private cloud (dedicated resources with elastic scaling), and Coolwei cloud (core data local, compute in the cloud).

The system is built for 7×24 operation. Health checks monitor runtime state and isolate faults automatically, services restart on failure, and load peaks trigger graceful degradation that protects core functions. QPS, latency and throughput are monitored live, alerts reach email, SMS and DingTalk, and performance reports are issued on a regular cycle. Multi-level access control and complete operation logs support security audit and incident tracing.

Where Yanlan is already running

The Yanlan model runs in production across the following high-standard content settings.