LLM//BOTTLENECK
2.8T total · 104B active parameters.
Moonshot's official card declares 2.8T total parameters and 104B active parameters per token. The U8 safetensors total is retained as source metadata but is not shown as a falsely precise logical count. Open the official parameter note ↗
Sized at 8,192 tokens of context with the whole model in device memory, across 18 common devices. Speeds are estimates, not benchmarks.
| Format | Size | Range | Basis |
|---|---|---|---|
| FP16 | 5602.4 GB | 3920.0 GB – 5656.1 GB | size range |
| Q8_0 | 2977.4 GB | 2082.5 GB – 3007.0 GB | size range |
| Q6_K | 2299.5 GB | 1607.8 GB – 2322.7 GB | size range |
| Q5_K_M | 2000.3 GB | 1347.5 GB – 2322.7 GB | size range |
| Q5_0 | 1954.0 GB | 1347.5 GB – 2322.7 GB | size range |
| Q4_K_M | 1718.8 GB | 1102.5 GB – 2322.7 GB | size range |
| Q4_0 | 1629.3 GB | 1102.5 GB – 2322.7 GB | size range |
| Q3_K_M | 1400.3 GB | 842.2 GB – 2322.7 GB | size range |
| Q2_K | 1108.8 GB | 643.1 GB – 2322.7 GB | size range |
A file size is not the memory a run needs: the KV cache and the runtime reserve come on top, and the calculator adds both.
| Device | Memory | Verdict | Decode | Calculator |
|---|---|---|---|---|
| NVIDIA GeForce RTX 5070 | 1720.4 GB of 12 GB | 93 layers on system RAM | <0.1 tok/s with offload | Open → |
| AMD Radeon™ RX 9070 XT | 1720.4 GB of 16 GB | 93 layers on system RAM | <0.1 tok/s with offload | Open → |
| NVIDIA GeForce RTX 4080 | 1720.4 GB of 16 GB | 93 layers on system RAM | <0.1 tok/s with offload | Open → |
| NVIDIA GeForce RTX 5080 | 1720.4 GB of 16 GB | 93 layers on system RAM | <0.1 tok/s with offload | Open → |
| NVIDIA GeForce RTX 5070 Ti | 1720.4 GB of 16 GB | 93 layers on system RAM | <0.1 tok/s with offload | Open → |
| NVIDIA GeForce RTX 5060 Ti 16GB | 1720.4 GB of 16 GB | 93 layers on system RAM | <0.1 tok/s with offload | Open → |
| NVIDIA GeForce RTX 4090 | 1720.4 GB of 24 GB | 92 layers on system RAM | <0.1 tok/s with offload | Open → |
| AMD Radeon™ RX 7900 XTX | 1720.4 GB of 24 GB | 92 layers on system RAM | <0.1 tok/s with offload | Open → |
| NVIDIA GeForce RTX 3090 | 1720.4 GB of 24 GB | 92 layers on system RAM | <0.1 tok/s with offload | Open → |
| NVIDIA GeForce RTX 5090 | 1720.4 GB of 32 GB | 92 layers on system RAM | <0.1 tok/s with offload | Open → |
| NVIDIA RTX 6000 Ada Generation | 1720.4 GB of 48 GB | 91 layers on system RAM | <0.1 tok/s with offload | Open → |
| Apple M4 Pro | 1720.4 GB of 64 GB | does not fit | not calibrated | Open → |
| Apple M5 Pro | 1720.4 GB of 64 GB | does not fit | not calibrated | Open → |
| AMD Ryzen AI Max+ 395 with Radeon 8060S | 1720.4 GB of 128 GB | does not fit | not calibrated | Open → |
| Apple M3 Max | 1720.4 GB of 128 GB | does not fit | not calibrated | Open → |
| Apple M4 Max | 1720.4 GB of 128 GB | does not fit | not calibrated | Open → |
| NVIDIA DGX Spark | 1720.4 GB of 128 GB | does not fit | not calibrated | Open → |
| Apple M2 Ultra | 1720.4 GB of 192 GB | does not fit | not calibrated | Open → |
Decode is a calibrated estimate from the published calibration; capacity is the manufacturer's published ceiling, not guaranteed free memory. Fitting in memory is not the same as loading: whether the runtime and version you have supports this architecture and format on that machine has not been tested here. Offloaded rows assume there is enough system RAM for the overflow — the calculator checks that against the RAM you declare. Each “Open” link keeps this model, format and context.
Every rentable machine that holds the whole model, from Vast.ai and RunPod, with this site's speed estimate for each and the provider's own price, read live. Cheapest first; sort by cost per token or speed instead, or switch the format.
The machines that hold Kimi-K3 at Q4_K_M are not in stock at Vast.ai or RunPod right now.
Nothing to set up and nothing to switch off: you pay only for the tokens you use.
Every machine holds the whole model at Q4_K_M and 8,192 tokens of context, no offload. Speeds are this site’s single-stream decode estimates; prices are what each provider’s own API quoted, on-demand, for the whole machine. Vast hosts below 98% measured reliability are left out.
Referral links Vast.ai, RunPod and Novita pay us a share of what you spend if you sign up through these buttons. It costs you nothing, and it never decides an order or a recommendation: both are computed from the live price and the speed, and options that pay us nothing are listed and recommended on the same terms. How we rank
Buying a card instead, or paying by the token? Compare a year of Kimi-K3 three ways →
The numbers every figure above is computed from, with the file they came from.
Retrieved 2026-09-01 at pinned commit a590ce090cb0; the native packed/fused checkpoint metadata is not used as a logical parameter total.
One page per device: whether Kimi-K3 fits, at which formats, and how fast.
curl -s https://llmbottleneck.com/v1/analyze \
-H "Authorization: Bearer $LLMB_KEY" -H "content-type: application/json" \
-d '{"model":"moonshotai-kimi-k3","quantization":"Q4_K_M","context":8192,"hardware":"nvidia-geforce-rtx-4090"}'Same engine, same evidence, every field sourced. Free key in one step →
<script src="https://llmbottleneck.com/widget.js" data-model="moonshotai-kimi-k3" data-quantization="Q4_K_M" data-hardware="nvidia-geforce-rtx-4090" data-context="8192"></script>
No key needed. Unbranded, with your own buy button, on Pro and Business →
LLM Bottleneck. “Kimi-K3 VRAM and hardware requirements.” Architecture from moonshotai/Kimi-K3 at revision a590ce090cb0, retrieved 2026-09-01. https://llmbottleneck.com/models/moonshotai-kimi-k3
Every figure above is either the published value or a reconstruction whose measured error is on the accuracy page.