Open‑Source AI Evolution: From Zipformer to Zapformer and Smart Automotive Quality

On December 10, 2025, the MEET 2026 Intelligent Future Conference in Beijing gathered more than twenty industry leaders, including Tsinghua’s Zhang Yaqin and AI Institute deputy Sun Maosong, to discuss how AI is crossing boundaries of industry, discipline, and scenario.

AI Evolution Analogy

Daniel Povey, chief voice scientist at Xiaomi, compared the decades‑long development of AI to biological evolution, arguing that AI progress relies on reproducible baseline systems that are iteratively refined, much like cellular division. He highlighted external constraints such as compute power and data quality, and noted that AI can also face bottlenecks analogous to the “Great Oxygenation Event,” but unlike random biological mutation, AI evolution is directed by human rationality and creativity.

Strategies for Continuous Breakthrough

Povey proposed several ways to keep AI advancing: strengthening mathematical foundations, fostering open‑source ecosystems, pursuing cross‑task and cross‑disciplinary integration, balancing general‑purpose and specialized models, and applying deeper understanding and smarter debugging.

Xiaomi’s AI Investment and Open‑Source Roadmap

Xiaomi plans to allocate 25 % of its 2025 R&D budget to AI, embedding AI across its high‑end product line and building an “AI‑for‑Science” platform that has already produced a titanium alloy supporting the SU7 super‑die‑casting technology.

From Zipformer to Zapformer

The company’s Kaldi‑based speech project, led by Povey (the “father of Kaldi”), introduced Zipformer in 2024, a speech encoder that set multiple international records through a novel sampling mechanism and the ScaledAdam optimizer. In 2025, Zapformer was released as a more forward‑looking universal sound foundation.

From “human voice” to “all sounds”: the model expanded from focusing on speech to handling environmental audio and other modalities.

From structural optimization to theoretical innovation: using Povey’s original Gradient Flow theory, Zapformer improves speech‑recognition accuracy by an additional 10‑15 % over Zipformer.

From task‑specific tuning to robust generalization: dropout layers were removed and the optimizer upgraded to TransformAdam, boosting large‑scale data fitting, convergence speed, and stability.

These advances collectively redefine the capabilities of a universal sound foundation.

Automotive Quality Multi‑Agent Engine

With the 50 000th vehicle rolling off Xiaomi’s Beijing factory, the company launched a full‑lifecycle quality engine powered by AI. The solution integrates data‑question answering, intelligent diagnosis, and one‑click 8D report generation, creating a “discover‑analyze‑solve‑prevent” loop that reduces defect resolution time from weeks to minutes.

The engine features a dual‑stage FAS/HAS agent architecture that supports multi‑turn dialogue, chart generation, and rapid analysis, compressing complex topic investigation from weeks to minutes.

By embedding massive Xiaomi domain data into an open‑source foundation model, the system links design, manufacturing, supply chain, sales, and after‑sales data into a unified knowledge graph, enabling root‑cause tracing across the entire vehicle lifecycle.

Natural‑language queries are translated to SQL via NL2SQL and intent recognition, raising complex query accuracy from 58 % to 86 % and eliminating “blind” data exploration.

Intelligent Diagnosis

The team introduced a hybrid “reasoning + temporal” large model that locates faults tenfold faster than traditional methods, achieving over 92 % anomaly detection on multi‑factor quality issues. The temporal base model also demonstrates leading performance with limited samples and can be continuously iterated to lower diagnostic costs.

Conclusion

From biological evolution to AI model breakthroughs, Xiaomi illustrates how deep understanding of foundational technology, open‑source collaboration, and cross‑industry application can reshape value creation. The transition from Zipformer to Zapformer, the AI‑driven titanium alloy, and the automotive quality engine together demonstrate AI’s capacity to penetrate industry boundaries and drive intelligent futures.