Inside Alibaba Cloud’s Tongyi Lingma: How Its Code Model Earned the Top 4+ Rating

On June 26, 2024, the China Academy of Information and Communications Technology announced the results of its Trusted AI Code Model Evaluation, awarding Alibaba Cloud’s Tongyi Lingma code model the top‑tier 4+ rating, indicating excellent performance in both general and specialized scenarios and a mature management framework.

Interview Overview

Q: Please introduce your company and the code model you submitted for evaluation.

A: The submitted model is Tongyi Lingma, built on the Tongyi foundation model with an expanded vocabulary and context length of 64K‑128K. It supports advanced code completion (including generation, long‑sequence modeling, modification, and SQL writing) and research‑question answering for tasks such as code conversion, optimization, unit‑test generation, and intelligent Q&A.

Q: How does it feel to receive the comprehensive 4+ rating, making you one of the first enterprises to achieve this level?

A: Since its November 2023 release, Tongyi Lingma has become the most downloaded plugin in China (over 3.5 million downloads). Enterprises are eager to adopt AI‑assisted programming to boost development efficiency. Continuous upgrades, such as the code‑focused codeqwen1.5 model and the research‑question answering qwen2 model, along with new features like retrieval‑enhanced code completion, fine‑tuning, and enterprise‑grade security, have pushed AI‑generated code to exceed 30% of total output.

Q: What motivated your participation in the evaluation, and what were the biggest takeaways?

A: The evaluation standard from the Academy is the national benchmark for code models. Participating and passing as a pioneer validates the model’s comprehensive abilities. The assessment covered thousands of test cases across general, specialized, and maturity dimensions, revealing many improvement opportunities despite the successful pass.

Q: What are the current highlights and core advantages of your code model?

A: Tongyi Lingma is recognized for its global‑leading capabilities. The codeqwen model uses a GQA architecture, trained on over 3 TB of code‑related tokens, supports hundreds of programming languages, and handles up to 64K context. Its performance in code generation, long‑sequence modeling, modification, and SQL tasks makes it the preferred AI coding assistant for many enterprises.

Q: Can you share typical enterprise applications and the efficiency gains they bring?

A: The core product is the Tongyi Lingma Intelligent Coding Assistant, which delivers a 10%‑15% overall productivity boost for companies. Additional explorations include applying the model across the full DevOps lifecycle—requirements management, test generation, deployment assistance—and future work on multi‑agent autonomous programming.

Q: What are your next steps for the code model?

A: Future releases will feature stronger capabilities, wider context windows, and more code‑related skills, targeting enhanced code retrieval, project‑level task solving, and code‑and‑test agents to meet complex engineering efficiency demands.

Q: How do you view the future of code‑model development?

A: Coding assistance is one of the biggest application areas for large models. As capabilities improve, new products will emerge, potentially disrupting software development within 1‑2 years. AI will increasingly replace routine engineering tasks, and the industry must embrace and contribute to AI‑native development paradigms.