Best GPU for Kimi K2: VRAM Requirements
Kimi K2 needs about 2,257.8 GB of GPU memory in BF16 and 564.9 GB at 4 bits. See which GPUs fit it for inference and fine-tuning.
Kimi K2 needs about 2,257.8 GB of VRAM in FP16, 1,129.2 GB in FP8 and 564.9 GB at 4 bits for inference with an 8,192-token context. In FP16 it needs more than eight cards of any GPU listed here; a 4-bit build fits on 4 × NVIDIA B200 180GB. Its weights are published in FP8, so the FP16 figure is for reference.
Kimi K2 is a mixture-of-experts language model with 1,026B total and 32B active parameters per token from Moonshot AI, released in July 2025, with a 131,072-token context window and published under the Modified MIT License. Its weights alone take 2,052 GB in BF16 and 513 GB at 4 bits per parameter; the tables below add the KV cache and runtime memory, then list the GPUs that hold each setup.
Recommended GPUs for Kimi K2
Best for inference
8 GPUs × NVIDIA B200 180GB
FP8 or INT8 weights, 1,129.2 GB at 8,192 tokens.
Best for fine-tuning
4 GPUs × NVIDIA B300 288GB
QLoRA on 4-bit weights, 666.9 GB.
Budget pick
8 GPUs × NVIDIA RTX PRO 6000 Blackwell 96GB
Smallest card class that holds a 4-bit build, 564.9 GB.
Kimi K2 VRAM requirements by precision
One request with an 8,192-token context, KV cache in BF16, plus a 10 % margin for activations and runtime buffers.
| Precision | Weights | KV cache | Total for inference |
|---|---|---|---|
| BF16 / FP16 | 2,052 GB | 0.6 GB | 2,257.8 GB |
| FP8 (published weights) | 1,026 GB | 0.6 GB | 1,129.2 GB |
| INT4 | 513 GB | 0.6 GB | 564.9 GB |
Kimi K2 VRAM by task
Fine-tuning rows assume a short sequence, batch 1 and gradient checkpointing; longer sequences add activation memory.
| Task | GPU memory |
|---|---|
| Inference at the full 131,072-token context (FP8) | 1,138.7 GB |
| QLoRA (4-bit base) | 666.9 GB |
| LoRA (BF16 base) | 2,667.6 GB |
| Full fine-tuning (BF16, AdamW) | 18,057.6 GB |
GPUs that fit Kimi K2
Number of cards whose combined memory holds each setup, assuming the model is split across them with tensor or pipeline parallelism. A dash means more than eight cards.
| GPU | Memory | BF16 inference (2,257.8 GB) | FP8 inference (1,129.2 GB) | INT4 inference (564.9 GB) | QLoRA (666.9 GB) | LoRA (2,667.6 GB) |
|---|---|---|---|---|---|---|
| NVIDIA A100 PCIe 80GB | 80 GB | — | — | 8 GPUs | — | — |
| NVIDIA A100 SXM 80GB | 80 GB | — | — | 8 GPUs | — | — |
| NVIDIA H100 PCIe 80GB | 80 GB | — | — | 8 GPUs | — | — |
| NVIDIA H100 SXM 80GB | 80 GB | — | — | 8 GPUs | — | — |
| NVIDIA RTX PRO 6000 Blackwell 96GB | 96 GB | — | — | 8 GPUs | 8 GPUs | — |
| NVIDIA H200 141GB | 141 GB | — | — | 8 GPUs | 8 GPUs | — |
| NVIDIA B200 180GB | 180 GB | — | 8 GPUs | 4 GPUs | 4 GPUs | — |
| NVIDIA B300 288GB | 288 GB | 8 GPUs | 4 GPUs | 2 GPUs | 4 GPUs | — |
| NVIDIA GB300 NVL | 288 GB | 8 GPUs | 4 GPUs | 2 GPUs | 4 GPUs | — |
What is Kimi K2?
Kimi K2 is a mixture-of-experts model from Moonshot AI with 384 experts, 8 selected per token plus one shared expert, so it computes 32 billion parameters per token out of about 1 trillion. Trained with the Muon optimizer, it targets coding, tool use and agentic tasks.
- Developer
- Moonshot AI
- Released
- July 2025
- Parameters
- 1,026B total, 32B active per token
- Context window
- 131,072 tokens
- License
- Modified MIT License
The model card rounds the total to 1 trillion; the published FP8 checkpoint holds 1.03 trillion parameters, the figure used here. Like DeepSeek V3, it uses multi-head latent attention, counted as a KV width of 288.
Source: Kimi K2 Instruct model card (as of November 17, 2026).
Frequently asked questions
- How much VRAM does Kimi K2 need?
- For inference with an 8,192-token context and one request, Kimi K2 needs about 2,257.8 GB in BF16, 1,129.2 GB in FP8, 564.9 GB in INT4, including a 10 % runtime margin. Real usage depends on the inference engine and its settings.
- Can Kimi K2 run on a single GPU?
- Not among the GPUs listed here: even a 4-bit build needs about 564.9 GB, so the model is split across several GPUs with tensor or pipeline parallelism.
- How much GPU memory does fine-tuning Kimi K2 take?
- QLoRA (4-bit base): about 666.9 GB; LoRA (BF16 base): about 2,667.6 GB; Full fine-tuning (BF16, AdamW): about 18,057.6 GB. Longer training sequences and larger batches add activation memory on top.
- What is the best GPU for Kimi K2?
- Best for inference: 8 GPUs × NVIDIA B200 180GB (FP8 or INT8 weights, 1,129.2 GB at 8,192 tokens). Best for fine-tuning: 4 GPUs × NVIDIA B300 288GB (QLoRA on 4-bit weights, 666.9 GB). Budget pick: 8 GPUs × NVIDIA RTX PRO 6000 Blackwell 96GB (Smallest card class that holds a 4-bit build, 564.9 GB). Check live offers before booking, as availability changes.
- Can Kimi K2 run on an RTX 4090?
- No: even a 4-bit build needs about 564.9 GB, more than eight RTX 4090 cards hold (192 GB). It needs data center GPUs with more memory per card.
- Can Kimi K2 run on an RTX 5090?
- No: even a 4-bit build needs about 564.9 GB, more than eight RTX 5090 cards hold (256 GB). It needs data center GPUs with more memory per card.
Estimates use the formulas of the LLM VRAM calculator (decimal GB). Real usage depends on the inference engine, quantization format and memory fragmentation: test on the target GPU before a long rental.
Explore the live market
Every offer on the live market comes from a reviewed supplier and is sold by EverXCloud Inc.