Kimi K3.1 VRAM requirements
Memory needed to run Kimi K3.1 at every published size and quant, with an 8k context. Each figure is weights plus KV cache plus runtime buffers, from the same memory model as the VRAM calculator. This family's attention design is not on file, so it is sized as plain grouped-query attention, which errs towards needing more.
Total memory by size and quant
GB at an 8k context. Bold is the Q4_K_M column, the usual starting point.
| Size | Q8_0 | Q6_K | Q5_K_M | Q4_K_M | Q3_K_M |
|---|---|---|---|---|---|
| 2800B | 2804 | 2279 | 1929 | 1579 | 1229 |
KV cache by context length
GB the context alone adds on top of the weights. Add this to a weights figure to size a longer session.
| Size | Weights at Q4_K_M | 4k ctx | 16k ctx | 32k ctx | 128k ctx |
|---|---|---|---|---|---|
| 2800B | 1575 | 1.1 | 4.3 | 8.6 | 34 |
What each size fits on at Q4_K_M
Hardware whose memory holds the model with headroom, smallest pool first. Fits means under 85% of the pool at an 8k context. A Mac's pool is what macOS lets the GPU use - two thirds of its memory up to 32 GB, three quarters from 36 GB - not the whole of it.
- 2800Bneeds about 1579 GB
Does not fit any single GPU or Mac on llamaperf at Q4_K_M. It needs a multi-GPU rig, a smaller quant, or CPU offloading.
Frequently asked
How much VRAM does Kimi K3.1 need?
About 1579 GB at Q4_K_M with an 8k context for the 2800B model, including the KV cache and runtime buffers. Q8 needs more and Q3 less; the table above lists every rung.
Can I run Kimi K3.1 on a 16GB GPU?
Not comfortably. Even the 2800B model needs about 1579 GB at Q4_K_M with an 8k context, which is more than a 16 GB card holds with headroom. A lower quant or CPU offloading can still get it running, at a cost in quality or speed.
Can I run Kimi K3.1 on a 24GB GPU such as an RTX 3090 or 4090?
Not comfortably. Even the 2800B model needs about 1579 GB at Q4_K_M with an 8k context, which is more than a 24 GB card holds with headroom.
How are these numbers calculated?
Weights at the quant's bits per weight, plus a KV cache sized from the model's attention design for the chosen context, plus compute buffers and a fixed runtime allowance. It is the same memory model the VRAM calculator uses. A family whose attention profile is not on file is sized as plain grouped-query attention, which errs towards needing more memory.