Rent a GPU

Can the NVIDIA RTX 5090 run Kimi K3.1 Pro 2.8T-A104B?

RTX 5090 × Kimi K3.1 Pro 2.8T-A104B8K context
NoDoes not fit in memory
Memory at Q4_K_M1698.3 GB of 32 GB
Weights 1697.5 GBKV cache 0.2 GBRuntime 0.6 GB
  1. FP16
  2. Q8_0
  3. Q6_K
  4. Q5_K_M
  5. Q4_K_M
  6. Q3_K_M
  7. Q2_K

No. Even at Q4_K_M, Kimi K3.1 Pro 2.8T-A104B needs about 1698.3 GB, but the NVIDIA RTX 5090 has 32 GB usable. You could offload some layers to system RAM, but it will be very slow.

Try other context lengths

Run it in the cloud

Too big for the NVIDIA RTX 5090? Run Kimi K3.1 Pro 2.8T-A104B on a rented GPU instead.

Rent a cloud GPU by the hour. No upfront cost, stop any time. Not sure whether to rent or buy? Compare the cost.

Rent a GPU on RunPodRent a GPU on Vast.ai

Memory needed by quantization

At 8,192 tokens of context. The black line marks the 32 GB the NVIDIA RTX 5090 can use.

QuantQualityMemoryFits?Speed
FP16Full quality5600.8 GBToo big—
Q8_0Near-lossless2975.8 GBToo big—
Q6_KExcellent2296.8 GBToo big—
Q5_K_MVery good1992.3 GBToo big—
Q4_K_MGood, most popular1698.3 GBToo big—
Q3_K_MNoticeably degraded1369.3 GBToo big—
Q2_KHeavily degraded1050.8 GBToo big—

Measured speed from real RTX 5090 owners

Tokens per second that people measured on their own NVIDIA RTX 5090, next to our estimate.

No one has measured Kimi K3.1 Pro 2.8T-A104B on the NVIDIA RTX 5090 yet. The speeds above are estimates. Run the benchmark and yours becomes the first real number on this page.

Have an RTX 5090? Add your real speed

One command times Kimi K3.1 Pro 2.8T-A104B in your own Ollama, shows you exactly what it would send, and asks before sending anything. No sign-up, no personal data.

macOS and Linux

curl -fsSL https://gpufit.dev/bench.sh | sh

Windows (PowerShell)

irm https://gpufit.dev/bench.ps1 | iex

How the benchmark works

Models that run well on the NVIDIA RTX 5090

Frequently asked questions

Can NVIDIA RTX 5090 run Kimi K3.1 Pro 2.8T-A104B?

No. Even at Q4_K_M, Kimi K3.1 Pro 2.8T-A104B needs about 1698.3 GB, but the NVIDIA RTX 5090 has 32 GB usable. You could offload some layers to system RAM, but it will be very slow.

How much VRAM does Kimi K3.1 Pro 2.8T-A104B need?

About 1698.3 GB at Q4_K_M with 8,192 tokens of context, and about 5600.8 GB at full FP16 precision.

How fast is Kimi K3.1 Pro 2.8T-A104B on the NVIDIA RTX 5090?

It does not fit in the NVIDIA RTX 5090's memory, so most layers would run from system RAM and generation would be very slow.