China’s Moonshot AI Sparks Global Buzz With ‘Kimi-K2’: “Smarter Than DeepSeek... at a Lower Cost”

Moonshot AI, a fast-rising Chinese startup, is rapidly expanding its presence in the global developer community with the release of its open-source model “Kimi-K2-Thinking.”

Since launch, the model has been dubbed “the second coming of DeepSeek,” and reports that it trained at a lower cost than DeepSeek have intensified attention.

“$1.1 Million Cheaper Than DeepSeek”… Trained on Older NVIDIA GPUs

According to CNBC, Moonshot AI’s latest model was trained for roughly $4.6 million, cheaper than DeepSeek-V3’s $5.57 million.

Moonshot did not disclose the exact figure, but during a Reddit AMA held on Nov. 10 (local time), a company representative stated that the model was trained “at a much lower cost than American companies.”

He added that the team used older NVIDIA H800 GPUs, saying, “Even hardware secured before the U.S. GPU sanctions was enough to achieve competitive performance.”

The account that made the statement matched the handle “ppwwyyww,” frequently used by Moonshot AI co-founder Yuxin Wu, leading many to interpret it as insider information.

‘Mooncake’ + ‘Linear Attention’: The Algorithmic Duo That Boosted Efficiency Fivefold

Moonshot AI attributed its cost cuts to two technical breakthroughs: the Mooncake training algorithm and its Kimi Linear Attention architecture.

Mooncake reportedly delivers up to five times higher training efficiency compared to conventional methods, while Linear Attention accelerates decoding speed and dramatically reduces memory usage.

As a result, Moonshot claims it achieved near–GPT-5-level performance without access to high-end GPUs.

Benchmark Results: Chasing GPT-5—And Aiming for the Open-Source Crown

According to the Intelligence Score released by benchmark firm Artificial Analysis, ▲GPT-5 / GPT-5 Codex: 68 ▲Kimi-K2-Thinking: 67

This puts Kimi-K2 just one point behind GPT-5 and ahead of models like Grok-4 (65) and Claude 4.5 Sonnet (63).

Notably, Kimi-K2 outperformed GPT-5 in BrowseComp (web search) and SEAL-0 (real-world information retrieval), strengthening its reputation as a practical, high-utility AI system.

“300 Steps of Autonomous Tool Use”—A Fully Realized AI Agent

Moonshot AI significantly upgraded its agent capabilities with this release.

Kimi-K2 can autonomously execute 200–300 consecutive tool calls, handling tasks such as search, code execution, and planning without human intervention.

Industry watchers say this marks a shift toward “thinking agents” capable of solving complex problems beyond the limits of traditional LLMs.

Explosive Community Response: “When Is K3 Coming?”

Since launch, Kimi-K2-Thinking has topped the open-source charts with over 80,000 downloads on Hugging Face.

Posts featuring the model on X (formerly Twitter) have surpassed 4.6 million views, and the global developer community is already asking, “When’s K3 coming?”

A Moonshot representative joked:

“Not until Altman finishes his trillion-dollar data center.”

The Return of Chinese AI: “K2 Is the New Challenger in Open Source”

Moonshot AI first entered the open-source arena in July with “Kimi-K2,” though it was overshadowed by models like GPT-4 and Grok-4 at the time.

But the new K2-Thinking release—praised for combining efficiency, performance, and practical utility—has pushed the company back into the global spotlight.

Tech analysts say the model signals both a “declaration of technological independence” for China’s AI industry and “the beginning of an open-source counterattack.”

By Ju-Baek Shinㅣjbshin@kmjournal.net