A workshop owner in a second-tier city hears that a small-business deduction changed this year. She asks a chatbot. It answers fluently — quoting a regulation that was repealed two years ago, with a document number that looks plausible and checks out nowhere. Nothing about the answer signals that it is stale. She books an appointment with an accountant she did not need, or worse, files on the old rule.
This is the failure mode that general-purpose large models (大模型) have never solved for high-stakes, fast-moving rulebooks, and it is the specific target of a small release from Beijing. In October, Beijing Lingjing Technology Co., Ltd. (北京灵境技术有限公司), working with the Beijing Huaxia Tax Agents Firm (北京华夏税务师事务所), released the Lingjing Tax Model (灵境税务大模型) with weights downloadable on ModelScope — which the developers describe as China's first open-source tax model.
Key takeaways
- The model has 8.95B parameters and is built on the Qwen3.5-9B-Base (通义千问基座) with full-parameter continued pretraining and fine-tuning, not a light adapter tune.
- Its corpus covers more than 8,000 national tax regulations plus accounting and audit rules, with every document tagged by validity status: in force, amended or repealed.
- Weights are free for individuals and research institutions; commercial use requires written authorisation, and the training code and corpus are not public.
- It runs on consumer hardware — a quantised build fits in well under 10 GB.
What a vertical model does differently
The pitch is not benchmark glory; the developers explicitly position it against the three ways general models fail on tax: citing repealed or amended rules, blurring computational boundaries, and producing conclusions that cannot be traced to a source. The engineering response happens mostly before training. The team assembled 11 corpus files totalling 48,867 raw documents and about 152.6 million characters, then deduplicated down to 45,868 documents and roughly 143 million characters. The collection layers more than 8,000 national tax regulations with accounting standards, audit standards and financial rules, 23,000 professional replies from the 12366 tax-service hotline, 2,999 court judgments and administrative enforcement cases, and 278 sample forms covering 12 types of tax filings — all sourced, the model card says, from official government websites.
The validity tagging is the quiet innovation. Because tax is a domain where last year's rule can be legally dangerous this year, every document carries an in-force, amended or repealed status, and answers are designed to surface that flag alongside the citation: regulation name, document number and article. Training followed in two full-parameter stages — 3,248 steps of continued pretraining on tax text, then 32,493 steps of instruction tuning on about 259,900 samples, of which roughly 228,000 are tax-specific. The training context was 4,096 tokens for pretraining and 8,192 for fine-tuning.
The listed capabilities are practical rather than flashy: statute question-answering with sources and validity hints; step-by-step tax computation with explicit boundaries (tax-inclusive versus tax-exclusive, threshold versus actual deductions, general versus simplified methods); case retrieval and reasoning; form-filling guidance for the 12 filing types; and consultation on audit and reconsideration procedures.
Why this template matters beyond tax
For a global audience, the interesting thing is not that a Chinese company shipped a domain model — it is the commercial shape of the release. Frontier labs sell general intelligence by subscription. This release suggests a different lane for open weights (开源权重): a small vertical model (垂域模型) whose value is a curated, maintained corpus rather than scale, distributed cheaply enough that the people who need it — bookkeepers, small-firm accountants, tax agencies — can actually run it.
The deployment numbers tell that story. The model ships as INT8 (about 10.99 GB) and a ladder of GGUF quantisations from roughly 4.62 GB to 9.53 GB, with a full-precision build at 17.9 GB — small enough for a consumer laptop, compatible with common local inference tooling. A compliance-aware vertical model that runs offline has an obvious audience in professions where client data cannot leave the building.
The release also stress-tests a governance pattern: weights open, corpus closed. Anyone can inspect the model's behaviour, but rebuilding or auditing the training data requires the vendor. No independent audit of the corpus exists, so its provenance rests on the developer's model card, and users are being asked to trust the methodology rather than verify it.
Honest limitations
- The evaluation is thin and self-reported: an internal 60-question set across six dimensions, with 85.6% coverage at full precision and 83.2% after 8-bit quantisation. No public benchmark comparisons exist.
- The developers' own model card warns that the model can fabricate document numbers, case numbers and amounts, and that every citation must be checked against original sources.
- The training context is short — 4K to 8K tokens — even though the architecture supports more, so long documents must be chunked carefully.
- The data cutoff is October 2026; tax rules issued after that are invisible to it, and the corpus requires ongoing maintenance that the open weights alone do not provide.
- It covers mainland China's tax system in simplified Chinese only, takes no tool calls or web searches, and its outputs explicitly do not constitute tax advice.
- Some corpus material was generated content rather than raw official text, which the developers themselves flag as needing sampling-based verification.
What readers can do now
- If you handle Chinese tax questions even occasionally, treat the model as a searchable, source-pointing reference — ask for the regulation name and document number, then read the original text before acting.
- If you build vertical models, borrow the methodology rather than the weights: corpus assembly, validity tagging and boundary-explicit computation are the transferable parts.
- If you run a small firm, the honest current use is preparation: let the model assemble the relevant rules and a draft calculation, then have a licensed professional confirm before anything is filed.