AI Research and Development in China: Models, Evaluation and Standards - Yenra

Start with the research question and the team

Each organization contains multiple projects, and collaborations cross institutional and national boundaries. These examples identify useful entry points. Follow the authors, affiliations and released artifacts for the particular result that interests you.

On a narrow screen, focus this table and use the arrow keys to scroll.

Use the map to choose a question: how reasoning is trained, how a model handles Chinese and other languages, whether a scientific task benefits from specialized data, or how an industrial trial measures success. An institution’s mission describes its intended work; individual studies establish particular results.

Follow the method behind a model release

The DeepSeek-R1 research report, first submitted in January 2025 and subsequently revised, investigates reinforcement learning for reasoning. In reinforcement learning, a training signal rewards selected outcomes or behavior. The R1 project distinguishes its RL-focused R1-Zero experiment from the additional training stages used for R1, and releases smaller distilled models trained using outputs from the larger model.

This gives a reader concrete comparisons to investigate: which training stages changed, how rewards were defined, and whether improvements hold on tasks beyond the training process. A distilled model has its own architecture and behavior; keep its exact identity attached to every result.

Alibaba’s April 2025 Qwen3 release illustrates another set of choices: dense and mixture-of-experts models, thinking and non-thinking modes, and multilingual support. A mixture-of-experts design activates selected parts of a model for each token. Comparing its total parameter count with another model’s active count mixes two definitions.

These dated releases are examples of research methods. For a current experiment, use the team’s release history and the exact model record. Record whether you have downloaded weights, a hosted API or a complete application. Weights enable some forms of inspection and reuse; reproducing training also depends on data, code, settings and computational resources. Check the terms attached to each artifact separately.

Look beyond chat to scientific and industrial work

The 2025 Intern-S1 paper describes a multimodal model trained on general and scientific material, with evaluations that include chemistry and materials tasks. Multimodal here means that the model can work with more than one kind of input representation. The research question is whether specialized training helps on demanding domain tasks while retaining broader capabilities.

The project’s repository now also lists Intern-S2 preview models, as checked on September 11, 2026. Keep the paper’s original model and benchmark results separate from those newer releases. A repository name can remain constant while its leading model, documentation and reported scores change.

Tsinghua AIR’s research-team descriptions include AI for science, with work on biological and chemical questions. Its university–industry structure offers a route for following how research meets practical requirements: start at a project, find the paper, then seek the relevant experiment or deployment study.

For a scientific claim, identify the measured endpoint. Predicting a molecular property on a held-out dataset, proposing a candidate material and confirming that material experimentally are distinct achievements. An industrial system likewise needs evidence from its actual workflow, including failures, human intervention and performance over time. The basic-research guide explains how those stages connect.

Make comparisons reproducible

OpenCompass provides evaluation tools and configurations for language models and, through its evolving integrations, multimodal tasks. Its public configuration files help reveal what a reported score actually measures. Save the code revision, dataset version, prompt format, scoring method and model settings used in a comparison.

Reasoning budgets, tool access and the number of attempts can change a result. For Chinese-language use, include the relevant writing style, terminology and task, alongside any English comparison. If prompts are translated, preserve the original and the translation and have a qualified reader check consequential differences.

Benchmark results are strongest when they can be independently reproduced and connected to the reader’s task. Preserve raw outputs, ambiguous scores and failed cases. The reproducible-research guide provides a broader framework for keeping that evidence together.

Connect standards to a defined measurement

China’s Ministry of Industry and Information Technology and three other bodies issued the 2024 national AI industry standardization-system guidelines. That document establishes a program of standards work. To investigate a specific requirement or test method, move from the program announcement to the individual standard’s official record.

For a concrete example, the official catalogue lists GB/T 45288.2-2025, on large-model testing and evaluation metrics and methods, as current, with publication and implementation dates of February 28, 2025. This catalogue status was checked on September 11, 2026. It identifies a published document; assessing a test against it requires reading the applicable text and examining the test evidence.

Keep three records distinct. A benchmark defines a task and scoring procedure. A standard supplies agreed specifications or methods within a stated scope. A deployment study measures how a system behaves in a particular environment. Each can contribute evidence, and each answers a different question.

When a report claims conformity, record the standard number, edition, scope, assessed system and supporting report. Check the official record for replacement or withdrawal and read the relevant provisions. A leaderboard position alone establishes no conformity result. Standards documentation and research performance also leave the applicable legal obligations to be assessed for the specific use.

Build a manageable reading routine

1. Choose one thread: for example, reasoning models, multilingual evaluation or materials prediction. Start with one of the primary projects above.
2. Capture the artifact: save the paper version, model identifier, release date and repository revision. Record what is actually available.
3. Write one supported result: state the task, comparison, conditions and uncertainty. Separate author-reported evidence from an independent reproduction.
4. Follow the next test: inspect a relevant failure case, reproduce a small suitable evaluation, or find the experiment needed for the proposed application.
5. Review changes: compare later releases against the saved record. Update the conclusion when methods, evidence or standard status change.

Use the AI research reading record (plain text) to keep the question, sources, comparison settings and next step together. The goal is a growing, checkable understanding of particular research contributions and their practical consequences.