Running Kimi K3 locally requires 1.56TB VRAM, limited to top-tier chips

ArtificialAnlys · x · 2026-07-31

Running Kimi K3 locally requires approximately 1.56 TB of memory for weights alone, meaning only NVIDIA B300 and AMD MI350X/MI355X single-node systems currently meet the bar, according to Artificial Analysis.

With 2.8 trillion parameters, it is the largest open-weights model ever released. Its hybrid attention mechanism reduces KV cache space, theoretically supporting over 2,000 concurrent requests at 256K context on a GB300 NVL72.

Utilizing the full 1M context window requires 1.59 TB of memory for a single user, growing by 30 GB per additional user. The total system cost reaches hundreds of thousands of dollars, and serving speed at high concurrency remains highly complex.

Original post →

More from Infra

Infra channel →