XNews

Kimi K3 Hardware Requirements and Benchmark Competition

Kimi K3 Hardware Requirements and Benchmark Competition

A user ran Kimi K3 on 80 Nvidia 5090 GPUs, achieving 20 tokens per second, a speed deemed usable. A visualization was created showing card requirements and bandwidth for various GPU platforms. Meanwhile, Claude Opus 5 with Max reasoning surpassed Kimi K3, taking first in Frontend Code Arena and Text Arena. The default high-reasoning version placed third in Code Arena (behind Kimi K3) and second in Text Arena.

Key insights

Practical Kimi K3 Deployment Requires 80 High-End GPUs

One user demonstrated that running Kimi K3 with 80 Nvidia 5090 GPUs achieves 20 tok/s, a performance level considered usable. A subsequent graphic, created based on this experiment and official data, illustrates the GPU count and bandwidth needed across different GPU platforms. The finding highlights the still-substantial hardware requirements for running such a large model.

Claude Opus 5 Overtakes Kimi K3 in Key Coding and Text Benchmarks

Claude Opus 5 with Max reasoning has claimed first place in both Frontend Code Arena and Text Arena, displacing Kimi K3 which had recently been dominating. The default high-inference version of Opus 5 also performed strongly, taking third in Code Arena behind Kimi K3 and second in Text Arena. This indicates a rapid reshuffling of performance leadership among large language models.

Source posts