What Do You Need to Run Kimi Locally? VRAM by Version
Kimi currently has 5 versions you can deploy locally. The smallest, Kimi Linear 48B A3B, needs 3GB of VRAM at Q4; the largest, Kimi K3, needs 33GB. The cheapest device that gets one running is the Arc Pro B65, with no used-price data yet.
VRAM needed for each Kimi version
5 versions| Version | Released | Context | ||||
|---|---|---|---|---|---|---|
| Kimi Linear 48B A3BMoE | 49.1B (3B active) | Oct 30, 2025 | 1M | 3plus 27 GB of RAM | Arc Pro B65— · 32 GB | 187,057 |
| Kimi K2MoE | 1,000B (32B active) | Sep 3, 2025 | 256K | 9plus 552 GB of RAM | GeForce RTX 2080 Ti— · 11 GB · Experts offloaded | 37,308 |
| Kimi K2.5MoE | 1,000B (32B active) | Jan 1, 2026 | 256K | 9plus 552 GB of RAM | GeForce RTX 2080 Ti— · 11 GB · Experts offloaded | 518,180 |
| Kimi K2.6MoE | 1,000B (32B active) | Apr 14, 2026 | 256K | 9plus 552 GB of RAM | GeForce RTX 2080 Ti— · 11 GB · Experts offloaded | 638,370 |
| Kimi K3MoE | 2,800B (104B active) | Jun 13, 2026 | 1M | 33plus 1,531 GB of RAM | A100 40GB PCIe— · 40 GB · Experts offloaded | 2,639,566 |
Minimum VRAM = Q4_K_M weights + KV cache for an 8K context + 1GB runtime overhead, rounded up. The cheapest device is the first one that fits when sorting by median used price on the current price basis.
About Kimi
Kimi is published by Moonshot AI under MIT / Modified MIT / Kimi K3 License. This page covers 5 versions ranging from 49.1B to 2,800B parameters; structure parameters and quantized sizes are synced weekly from Hugging Face.
VendorMoonshot AI
LicenseMIT / Modified MIT / Kimi K3 License
Official sitehttps://www.moonshot.ai/
Hugging Face orgmoonshotai