Moonshot · 1,026B parameters · 32B active
Kimi K2: GPU memory requirements and which GPU to rent
Moonshot AI's Kimi K2 puts about a trillion parameters — 32 billion active per token — on DeepSeek-V3's architecture, with the same compact latent-attention cache. At FP8 its weights alone pass a terabyte.
- Weights at FP81,026 GB2,052 GB at 16-bit · 564 GB at 4-bit
- FP8, 8k context, 1 sequence1,069 GBweights + KV cache + runtime
- Suggested here8× B200$18,900 a month · dedicated
Inference
How much GPU memory Kimi K2 needs to serve
Total memory for one sequence at each context length, and for 16 concurrent sequences at 8k — weights, KV cache at 16-bit and runtime overhead included.
| Weights precision | Weights | 4k context | 32k context | 128k context | 8k × 16 sequences | Suggested here, 8k |
|---|---|---|---|---|---|---|
| 16-bit (BF16 / FP16) | 2,052 GB | 2,136 GB | 2,138 GB | 2,145 GB | 2,145 GB | Multi-node |
| 8-bit (FP8 / INT8) | 1,026 GB | 1,069 GB | 1,071 GB | 1,078 GB | 1,078 GB | 8× B200 · $18,900/mo |
| 4-bit (GPTQ, AWQ, Q4) | 564 GB | 589 GB | 591 GB | 598 GB | 598 GB | 8× A100 · $5,192/mo |
GB are 109 bytes. Context lengths are shown for scale where the model supports them; an FP8 KV cache halves the cache part of each figure. As a mixture of experts, every expert stays resident, so memory follows the total parameter count, not the active one. Same latent-attention cache as DeepSeek-V3.
Hardware
GPU servers that run Kimi K2
For FP8 inference with an 8k context and one sequence: the configurations we would suggest, in order — one large card ahead of several small ones.
Fine-tuning Kimi K2
Memory for a 2k-token sequence and a micro-batch of 4, with gradient checkpointing. Full fine-tuning keeps gradients and Adam states for every parameter; LoRA and QLoRA train small adapters on a frozen base.
| Method | Memory | Suggested here |
|---|---|---|
| QLoRA, 4-bit base | 654 GB | 4× B200 · $10,760/mo |
| LoRA, 16-bit base | 2,201 GB | Multi-node |
| Full fine-tune, mixed precision | 16,514 GB | Multi-node |
Architecture at a glance
Same latent-attention cache as DeepSeek-V3. Figures from the model's published configuration; method on the calculator page.
Kimi K2 hardware: questions answered
How much VRAM does Kimi K2 need?
For the weights alone: about 2,052 GB at 16-bit, 1,026 GB at FP8 and 564 GB at 4-bit. Serving it at FP8 with an 8k-token context and one sequence takes about 1,069 GB in total once the KV cache and the runtime are counted; long contexts and many concurrent users add the KV cache on top — about 0.07 GB per 1,000 tokens at 16-bit.
Can Kimi K2 run on a single GPU?
Not at FP8: the weights need more memory than any single GPU in our catalogue holds. otherwise the suggested setup is 8× B200 · $18,900/mo.
What is the cheapest GPU setup to run Kimi K2?
For FP8 inference with an 8k context and one sequence, the configuration we would suggest is 8× B200 · $18,900/mo; at 4-bit, 8× A100 · $5,192/mo. Suggestions favour one large card over several small ones: splitting a model across PCIe cards costs speed.
How much GPU memory do I need to fine-tune Kimi K2?
With a 2k-token sequence and a micro-batch of 4: about 654 GB for QLoRA on a 4-bit base, 2,201 GB for LoRA on a 16-bit base, and 16,514 GB for a full fine-tune with Adam in mixed precision. Gradient checkpointing is assumed; sharding the optimizer across GPUs spreads the full fine-tune over several cards.
Run Kimi K2 on 8× B200 for $18,900 a month
Dedicated, single-tenant and live in under 10 minutes — with vLLM, the Hugging Face stack or Ollama preinstalled if you want. Paid in BTC, ETH or USDT, with no KYC.