How much memory does Kimi K3 need?
About 1,700 GB at Q4_K_M with an 8K context. That is more than any single Mac.
Try other settings in the calculatorMemory by quantization and context
Total memory: weights, KV cache in F16 and about 1 GB of buffers. Under each total, the smallest Mac that can run it with the default GPU memory limit.
| Quantization | 8K context | 32K context | 128K context |
|---|---|---|---|
| Q3_K_M | 1,360 GBtoo big for a Mac | 1,361 GBtoo big for a Mac | 1,363 GBtoo big for a Mac |
| Q4_K_M | 1,700 GBtoo big for a Mac | 1,701 GBtoo big for a Mac | 1,704 GBtoo big for a Mac |
| Q5_K_M | 1,982 GBtoo big for a Mac | 1,983 GBtoo big for a Mac | 1,985 GBtoo big for a Mac |
| Q6_K | 2,281 GBtoo big for a Mac | 2,281 GBtoo big for a Mac | 2,284 GBtoo big for a Mac |
| Q8_0 | 2,955 GBtoo big for a Mac | 2,956 GBtoo big for a Mac | 2,958 GBtoo big for a Mac |
| F16 | 5,561 GBtoo big for a Mac | 5,562 GBtoo big for a Mac | 5,564 GBtoo big for a Mac |
About the model
- Parameters
- 2,780B
- Type
- Mixture of experts, 896 experts
- Attention
- hybrid: linear attention plus compressed latent attention
- KV cache, F16
- 0.2 GB at 8K, 3.6 GB at 128K
- On Hugging Face
- moonshotai/Kimi-K3
Most layers use linear attention with a fixed-size state; the rest cache one compressed latent per token. The KV cache stays small even at long context. As a mixture-of-experts model it loads every expert into memory, so memory follows the total parameter count, while speed follows the smaller number of active parameters.
Questions
How much memory does Kimi K3 need?
About 1,700 GB at Q4_K_M with an 8K context: 1,699 GB of weights, 0.2 GB of KV cache and about 1 GB of buffers. At 32K context it needs about 1,701 GB.
Can Kimi K3 run on a Mac?
Not on a single Mac at Q4_K_M: it needs about 1,700 GB, more than the largest Mac memory size. It needs a cluster of machines or a lower-bit quantization.
How big is the Kimi K3 KV cache?
About 0.2 GB at 8K tokens and 3.6 GB at 128K tokens with an F16 cache. Its attention design: hybrid: linear attention plus compressed latent attention.
Other models
- Gemma 4 12B
- gpt-oss 20B
- Gemma 4 26B-A4B
- Qwen3.8 27B
- GLM-4.7 Flash
- Gemma 4 31B
- Qwen3.6 35B-A3B
- gpt-oss 120B
- Mistral Medium 3.5
- Qwen3.8 Flash Next
- DeepSeek V4 Flash
- GLM-5.3
- DeepSeek V4 Pro
Data read from Hugging Face on 2026-11-09. These are estimates; benchmarks on this site report measured memory. How the estimate works. Compare all models in the model fit chart.
Get the next one
New benchmarks and agent runs by email, with the raw data. Or follow on RSS or X.