Inside the 2025 AI+ R&D Survey: How Chinese Teams Are Transforming Software Development with Large Models

Industry Landscape Reshaped: From Model Competition to Ecosystem Competition

When large language models (LLMs) become a standard tool in software development, the industry shifts from tentative trials to deep integration. The survey shows 89.2% of teams have embraced LLMs, and 62.8% are actively applying them, an 8.8% increase over the previous year.

1. Application Depth Increases Significantly

Teams that were merely exploring LLMs dropped from 54% to 26.4%, with most converting to active users, indicating tangible benefits. Teams still observing remain at 7.4%, suggesting market education is largely complete.

2. Vertical Solutions Become Mainstream

Five sectors—finance, manufacturing, energy, software development, and government—now have mature industry models, expanding market size by ~1.8× compared to 2024. Software and IT services achieve over 78% penetration, with breakthroughs in lifecycle management, code intelligence, and documentation.

3. Model Selection Landscape Shifts Dramatically

DeepSeek surged from 3.2% to 81.1% to become the domestic leader, driven by its open‑source R1 and V3 releases. Alibaba Tongyi Qianwen holds 65.5%, Doubao rises to 41.9% (2.38×), while OpenAI GPT retains 67.6% globally. Claude series grows fastest to 43.2% (2.45×). The shift reflects the advantage of open‑source performance and the trend of using multiple models for diverse scenarios.

Core Value Realization: Cost Reduction and Efficiency Gains

85.1% of teams aim to achieve "cost reduction and efficiency improvement" with LLMs, a 10% rise year‑over‑year. Expectations for development efficiency rose 16.7% to 81.8%.

More than 70% of enterprises quantify LLM benefits via time savings, cost cuts, and quality gains. 33.1% report 20‑39% efficiency gains, 31.8% report 10‑20%, totaling over 64% of teams. Assuming a 20% average gain, a large enterprise with 10,000 developers could save about ¥1 billion annually.

LLMs reduce development and testing labor costs by 30‑45%, with 81.1% of teams using them for code generation—a 15% increase. Some large firms report 70% of code now generated by AI.

Full‑Lifecycle Penetration: From Local Breakthroughs to Systemic Integration

1. Requirements

"Refining or optimizing requirement documents" accounts for 60.8% of use cases, up from 47.6% in 2024 and 22.7% in 2023. "Reviewing requirements" and "extracting key points" rise to 47.3% and 46.6% respectively, while non‑use drops to 8.1%.

43.2% of firms see efficiency gains below 20%, 26.4% see 20‑40% gains, and uncertainty falls from 39.7% to 14.9%.

2. Design

In 2025, "consulting LLMs for design suggestions" leads with 61.5% usage, surpassing 48.4% in 2024. Knowledge‑base queries reach 56.1%, doubling from 2023. Other design assists such as "functional decomposition" (36.5%) and "UI/UX assistance" (35.1%) also grow.

Non‑use drops to 5.4% in 2025.

3. Programming

"Code completion" dominates at 70.9%, with "code optimization" at 60.8% and "function‑level generation" at 56.1%. The "not started" segment falls to 3.4%.

Tool adoption: Cursor leads with 18.9%, followed by Tongyi Lingma (16.9%) and in‑house tools (14.9%). Code adoption rates cluster at 25% for the 21‑30% range, with 47.2% of teams achieving ≥30% adoption.

4. Testing

"Test case generation" is the top scenario at 68.2%, with "test script generation" at 51.4% and "test data generation" at 31.1%. Non‑use declines to 16.9%.

Commercial tools gain traction; Ant TestAgent holds 6.8% market share. Efficiency gains: 23% of teams see 20‑39% improvement, another 23.6% see 10‑20%.

5. Operations

LLMs now cover "log analysis" (45.3%), "anomaly diagnosis" (43.9%), and "issue localization" (41.2%), each growing severalfold since 2023. 50.7% of firms remain uncertain about efficiency gains, down from 57.9%.

Technical and Organizational Dual Transformation

1. Tech Stack Upgrade

About 62% of enterprises use RAG or LoRA fine‑tuning to combine internal data with models, up from 30% in 2024. Open‑source model deployments exceed 50% (54.8%), with private deployment becoming mainstream for security and customization.

Agent frameworks see diversification: Dify leads with 36.5%, followed by LangChain (20.9%) and Coze (17.6%). Non‑use drops to 19.6%.

2. Organizational Change

New roles such as "Prompt Engineer", "Context Engineer", and "AI Agent Orchestrator" emerge. 50.7% of teams report faster personal learning of new technologies, 40.5% see junior developers handling complex tasks sooner, and 39.2% allocate more time to prompt design.

3. Governance

Governance expands from model security to full‑stack intelligent governance, covering data privacy, prompt rationality, model provenance, and controllable outputs. Domestic firms lead in security measures.

Metrics evolve: 40.5% use development efficiency indicators, 37.2% track quality metrics, and 33.8% estimate labor cost, moving from qualitative to quantitative assessment.

Challenges and Responses: Eight Major Difficulties

Lack of high‑quality training data (50.8%)

Security and privacy concerns (40.5%)

AI hallucinations causing hidden bugs (38.9%)

Unpredictable and poorly controllable AI Agent behavior (29.7%)

Insufficient application skill among R&D staff (28.4%)

Shortage of ML and LLM training talent (27.7%)

Insufficient compute resources or cloud platforms (27.7%)

Difficulty measuring ROI accurately (25%)

While data quality, talent, and compute are improving, safety, privacy, and hallucination remain core bottlenecks.

Future Outlook: Towards an Agentic Era

61.5% of respondents believe "everyone will have an AI assistant" and 58.8% expect AI to dramatically boost R&D efficiency.

Enterprises plan to introduce task‑oriented LLM agents by 2026 for cross‑stage automation and decision‑making, shifting from passive response to autonomous workflow execution.

Collaboration will evolve from "prompt‑reply" to "continuous dialogue and co‑creation," with LLMs becoming long‑term interactive members of product, development, and testing teams.

Conclusion: Embrace Change, Reshape R&D

The report shows that AI is not just a technological upgrade but a fundamental shift in industry mindset and organization, moving from pilot projects to systematic, vertical, and ecosystem‑driven adoption, heralding a 3.0 era of model‑driven software engineering.