Jensen Huang Showcases DeepSeek, Kimi AI Models for Next-Gen Chip Performance

During his keynote address at CES 2026, Nvidia CEO Jensen Huang featured Chinese large language models, including DeepSeek V3.2 and Kimi K2, to demonstrate the capabilities of the company's upcoming Rubin architecture. The presentation highlighted these models alongside Qwen as top-tier open-source AI, with performance levels approaching those of closed-source models. OpenAI's GPT-OSS and Nvidia's Nemotron were also mentioned.

Abstract visualization of a Mixture of Experts (MoE) neural network, showing selective activation of pathways.

DeepSeek-R1, Qwen3, and Kimi K2 were presented as examples of the Mixture of Experts (MoE) approach, which activates a limited number of parameters to reduce computational load and HBM memory bandwidth requirements.

Rubin Architecture Demonstrations

Huang specifically used DeepSeek and Kimi K2 Thinking to illustrate performance gains with the Rubin architecture. The integration reportedly increased Kimi K2 Thinking's inference throughput tenfold while reducing token cost to one-tenth of its original value.

Abstract visualization of rapidly increasing computing demand, with data streams and light trails expanding.

The presentation also included a slide on surging computing demand, featuring the 480B Qwen3 and 1TB Kimi K2 as representative models, indicating an annual order-of-magnitude increase in parameter scale.

Chinese Models as Performance Benchmarks

In a December Nvidia blog post, DeepSeek R1 and Kimi K2 Thinking were used for performance evaluations. Tests showed Kimi K2 Thinking's performance on the GB200 NVL72 could increase by a factor of 10. Separately, in the SemiAnalysis InferenceMax test, DeepSeek-R1 reportedly reduced the cost per million tokens by over 10 times, with Mistral Large 3 also achieving a tenfold acceleration. These results suggest that complex "thinking-type" MoE deployments are becoming feasible for daily applications.

Digital interface displaying AI performance metrics, with an upward-trending graph showing significant improvement.

The MoE architecture has become a mainstream choice in AI development, with statistics indicating that over 60% of open-source AI has adopted it since 2025. This architecture has contributed to a nearly 70-fold improvement in LLM intelligence levels since early 2023. The Artificial Analysis (AA) leaderboard shows that all top 10 most intelligent open-source models utilize the MoE structure. Nvidia's GB200 NVL72 is designed to support the deployment of such large MoE models.

Global Impact of Chinese AI

Late last year, Anthropic conducted a behavioral benchmark test across 16 global models. DeepSeek and Kimi were the only two Chinese entrants. Kimi K2 Thinking was noted for its low susceptibility to being misled, earning recognition as the "best-performing non-US model."

Marc Andreessen, a Silicon Valley venture capitalist, has publicly acknowledged Chinese AI. Additionally, OpenAI's former CTO's new product, Thinker, announced its integration with Kimi K2 Thinking last month.

Abstract visualization of global AI influence, with a bright, expanding cluster of interconnected nodes over Asia.

In the "2025 Open Source Model Review" by AI experts Nathan Lambert and Florian Brand, DeepSeek, Qwen, and Kimi secured the top three positions. Lambert further highlighted several advantages of Chinese open-source AI:

• Rapid Development: Chinese laboratories are releasing models at an accelerated pace, narrowing the gap with closed-source models.

• Focus on Usability: There is a shift from achieving high benchmark scores to improving practical usability, as seen in the evolution of models like Qwen. K2 Thinking's native 4-bit precision in post-training aims to support long-sequence RL expansion more efficiently for service tasks.

• Brand Recognition: DeepSeek, Qwen, and Kimi are gaining international recognition as representatives of Eastern technological strength.

• Advanced Capabilities: Kimi K2 Thinking's support for "hundreds of stable tool calls" and "interleaved thinking" (where the model thinks between tool calls) demonstrates advanced logical chain maturation.

• Competitive Pressure: The growth of open-source AI, particularly from China, is increasing pressure on American closed-source laboratories, challenging their value proposition based solely on benchmark scores.

Jensen Huang's CES 2026 speech emphasized open-source AI as a central theme, with Chinese open-source models demonstrating capabilities that are attracting global attention.