Why Kimi K2 Overtook DeepSeek to Become the Top Open‑Source AI Model

Why Kimi K2 Overtook DeepSeek to Become the Top Open‑Source AI Model

Thousands of users voted, and Kimi K2 surpassed DeepSeek to claim the global open‑source #1 position. International netizens flooded the comments with praise.

Today the arena updated Kimi K2’s ranking: open‑source #1, overall #5, closely trailing elite closed‑source models like Grok 4.

Open‑source first, overall fifth, and closing the gap with top closed models.

Kimi K2 also performs well in individual capabilities, matching or surpassing leading closed models:

Continuous multi‑turn conversation – ranked first, ahead of o3 and Grok 4 (both fourth).

Programming ability – second, on par with GPT 4.5 and Grok 4.

Complex prompt handling – second, alongside o3 and 4o.

Only two open‑source models entered the overall Top 10, both from China.

In the past week Kimi K2 attracted massive attention: 5.6 K GitHub stars and nearly 100 K Hugging Face downloads, not counting its use in Chinese communities.

Perplexity CEO personally endorsed Kimi K2, stating that internal evaluations showed excellent performance and that Perplexity plans further fine‑tuning based on K2.

Kimi K2 performed outstandingly in internal assessments; Perplexity will conduct subsequent post‑training on the K2 model.

High traffic and the model’s large size caused API slowdown, prompting an official announcement from Kimi.

Heavy traffic and large model size led to slow API responses.

There were doubts about whether Kimi K2 used DeepSeek V3 architecture. The Kimi team confirmed that it inherits the DeepSeek V3 architecture with additional parameter adjustments.

Indeed, Kimi K2 inherits DeepSeek V3’s architecture, with further parameter tuning.

The team considered alternative architectures but chose to fully adopt V3 because it is proven, cost‑effective, and fits their limited training and inference resources.

Four specific structural changes were made:

Increased number of experts – more MoE parameters improve loss without increasing active parameters.

Halved attention‑head count – saved cost offsetting larger MoE parameters with minimal impact.

Retained only the first dense layer – subsequent layers use MoE, barely affecting inference.

No expert grouping – free routing with dynamic re‑ordering (EPLB) handles load imbalance and enhances flexibility.

The resulting inference scheme, despite a 1.5× increase in total parameters, reduces theoretical prefill and decode latency (excluding communication), and does not significantly raise cost compared to V3.

Although total parameters grew 1.5×, prefilling and decoding times are smaller; even accounting for communication overlap, the scheme does not incur notable cost increases over V3.

This “cost‑effective” structural tuning has been described by Chinese netizens as “quite bold.”

Open‑Source Catching Up or Surpassing Closed‑Source

The stereotype that “open‑source means weaker performance” is being shattered; open‑source models are increasingly competitive.

Overall rankings are rising, and score gaps are narrowing. The top‑10 models all score above 1400, placing open‑source and closed‑source models on nearly the same starting line.

Kimi K2’s total score now closely approaches top closed models like Grok 4 and GPT 4.5.

Tim Dettmers, research scientist at the Allen Institute for AI, noted that open‑source beating closed‑source will become increasingly common.

Open‑source defeating closed‑source will become more frequent.

Perplexity’s CEO also emphasized that open‑source models will play a crucial role in the global diffusion of AI capabilities, enabling localized and customized AI experiences.

Open‑source models are vital for tailoring and localizing AI experiences worldwide.

In the top‑10, only two open‑source models remain, both domestic, highlighting the rising prominence of Chinese contributions.