Key Facts

• Who: China’s regulatory authorities governing large language models
• What: Three-tier risk classification system with quantitative compliance thresholds
• When: Filing system effective 2026
• Impact: Applies to all LLMs deployed in China regardless of origin

TL;DR

China has implemented a three-tier risk classification system for large language model filing, requiring high-risk models to undergo expert panel review lasting up to 3 months. The system establishes quantitative thresholds including a 95% refusal rate for sensitive content and a 30% maximum for overseas training data, affecting all LLMs deployed in China.

What Changed

China’s 2026 LLM filing system introduces a structured regulatory framework categorizing models into three risk tiers: standard, medium, and high-risk. The system, detailed in official guidelines, applies to all large language models deployed within China regardless of their country of origin.

The classification determines both the review process and the timeline for regulatory approval:

• Standard-risk models: Streamlined filing process with automated review
• Medium-risk models: Enhanced documentation requirements with departmental review
• High-risk models: Expert panel review requiring up to 3 months

According to the CSDN report, the filing requirements include two critical quantitative thresholds that will determine a model’s compliance status.

Why It Matters

The regulatory framework establishes specific, measurable criteria for LLM compliance in China:

Quantitative Thresholds

Metric	Threshold	Scope
Content safety refusal rate	≥ 95%	Sensitive query detection
Overseas training data	≤ 30%	Training dataset composition

The 95% refusal rate threshold requires models to accurately identify and refuse responses to sensitive content categories. This metric demands sophisticated content moderation systems with high precision and recall for Chinese-language content boundaries.

The 30% cap on overseas training data introduces supply chain constraints for model developers. International LLM providers relying on globally-sourced datasets face restructuring requirements or potential market exclusion.

Review Timeline Impact

Risk Level	Review Duration	Review Body
Standard	Days to weeks	Automated/Departmental
Medium	Weeks to 1 month	Departmental
High	Up to 3 months	Expert panel

The extended timeline for high-risk models creates go-to-market delays that affect commercial planning. Models classified as high-risk face a quarter-year approval window, requiring advance regulatory strategy.

🔺 Scout Intel: What Others Missed

Confidence: medium | Novelty Score: 68/100

The filing system’s quantitative thresholds create a regulatory asymmetry that favors domestic Chinese model developers. The 30% overseas training data cap disproportionately impacts foreign LLM providers who typically train on globally-sourced corpora—OpenAI’s GPT models, Anthropic’s Claude, and Google’s Gemini all derive significant training data from English-language sources. Meanwhile, Chinese domestic models like Baidu’s ERNIE and Alibaba’s Qwen already operate with predominantly domestic data, positioning them for smoother compliance.

The 95% refusal rate threshold requires culturally-specific content moderation infrastructure that cannot be imported from Western safety systems. Models trained primarily on non-Chinese data lack the linguistic and cultural context to achieve this threshold, creating a technical barrier that compounds the data composition requirement.

Key Implication: Foreign LLM providers seeking China market access face a dual compliance burden—dataset restructuring to meet the 30% cap, plus custom content moderation development for the 95% refusal threshold—that effectively necessitates a China-specific model variant, raising operational costs and delaying deployment timelines.

What This Means

For LLM Developers Entering China

Companies planning to deploy LLMs in China must evaluate their risk classification early. Models handling sensitive domains—healthcare, finance, legal, education—face higher likelihood of medium or high-risk designation. Pre-filing assessment should prioritize:

1. Dataset audit: Quantify overseas training data percentage and plan remediation if exceeding 30%
2. Content safety testing: Benchmark refusal rates against the 95% threshold using Chinese-language test sets
3. Timeline planning: Factor potential 3-month review delay into commercial launch schedules

For Domestic Chinese AI Companies

The regulatory framework provides clarity that enables investment decisions. Companies with domestic training data pipelines and Chinese-language content moderation systems gain compliance advantages. This may accelerate the divergence between China-market LLMs and global models, reinforcing the development of parallel AI ecosystems.

What to Watch

• Enforcement patterns: How regulators apply risk classifications in practice, particularly borderline cases
• Foreign provider responses: Whether international AI companies develop China-specific model variants or exit the market
• Threshold adjustments: Potential future changes to the 95% refusal rate or 30% overseas data caps based on industry feedback

Sources

• China LLM Filing System 2026 — CSDN, May 2026