LLMBOTTLENECK.COM
moonshotai / kimi_k2

Kimi-K2-Thinking

1T total · 32B active parameters, routing 8 of 384 experts per token.

Moonshot's official card declares 1T total parameters and 32B active parameters per token. The I32 safetensors total is a packed storage representation and is not shown as a falsely precise logical count. Open the official parameter note ↗

Open in the calculator →

Architecturepublished data
Quick answer
Kimi-K2-Thinking needs about 615.3 GB at Q4_K_M — the format most people download — with 8,192 tokens of context. None of the 18 common devices listed below holds it entirely at this setting. No hardware for it? Rent a GPU that runs it, priced live.

Sized at 8,192 tokens of context with the whole model in device memory, across 18 common devices. Speeds are estimates, not benchmarks.

Weight-file size by format

FormatSizeRangeBasis
FP162000.9 GB1400.0 GB – 2020.0 GBsize range
Q8_01063.4 GB743.8 GB – 1075.4 GBsize range
Q6_K821.3 GB574.2 GB – 831.3 GBsize range
Q5_K_M714.4 GB481.3 GB – 831.3 GBsize range
Q5_0697.9 GB481.3 GB – 831.3 GBsize range
Q4_K_M613.9 GB393.8 GB – 831.3 GBsize range
Q4_0581.9 GB393.8 GB – 831.3 GBsize range
Q3_K_M500.1 GB300.8 GB – 831.3 GBsize range
Q2_K396.0 GB229.7 GB – 831.3 GBsize 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.

What the labels mean
published data
Read from a published source — a file's byte count, a model's configuration or a manufacturer's specification — or exact arithmetic on such values. Not a measurement on a machine.
reconstructed size
Weight size reconstructed from the pinned architecture, because no published file exists.
size range
Only a lower and an upper bound are claimed for this weight size.

Hardware ladder

Q4_K_M at 8,192 tokens · 0 of 18 devices hold it entirely
DeviceMemoryVerdictDecodeCalculator
NVIDIA GeForce RTX 5070615.3 GB of 12 GB60 layers on system RAM~3.3 tok/s with offloadOpen →
AMD Radeon™ RX 9070 XT615.3 GB of 16 GB60 layers on system RAM~3.3 tok/s with offloadOpen →
NVIDIA GeForce RTX 4080615.3 GB of 16 GB60 layers on system RAM~3.3 tok/s with offloadOpen →
NVIDIA GeForce RTX 5080615.3 GB of 16 GB60 layers on system RAM~3.3 tok/s with offloadOpen →
NVIDIA GeForce RTX 5070 Ti615.3 GB of 16 GB60 layers on system RAM~3.3 tok/s with offloadOpen →
NVIDIA GeForce RTX 5060 Ti 16GB615.3 GB of 16 GB60 layers on system RAM~3.3 tok/s with offloadOpen →
NVIDIA GeForce RTX 4090615.3 GB of 24 GB59 layers on system RAM~3.4 tok/s with offloadOpen →
AMD Radeon™ RX 7900 XTX615.3 GB of 24 GB59 layers on system RAM~3.4 tok/s with offloadOpen →
NVIDIA GeForce RTX 3090615.3 GB of 24 GB59 layers on system RAM~3.4 tok/s with offloadOpen →
NVIDIA GeForce RTX 5090615.3 GB of 32 GB58 layers on system RAM~3.5 tok/s with offloadOpen →
NVIDIA RTX 6000 Ada Generation615.3 GB of 48 GB57 layers on system RAM~3.5 tok/s with offloadOpen →
Apple M4 Pro615.3 GB of 64 GBdoes not fitnot calibratedOpen →
Apple M5 Pro615.3 GB of 64 GBdoes not fitnot calibratedOpen →
AMD Ryzen AI Max+ 395 with Radeon 8060S615.3 GB of 128 GBdoes not fitnot calibratedOpen →
Apple M3 Max615.3 GB of 128 GBdoes not fitnot calibratedOpen →
Apple M4 Max615.3 GB of 128 GBdoes not fitnot calibratedOpen →
NVIDIA DGX Spark615.3 GB of 128 GBdoes not fitnot calibratedOpen →
Apple M2 Ultra615.3 GB of 192 GBdoes not fitnot calibratedOpen →

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.

Run it in the cloud

Rent a GPU that runs Kimi-K2-Thinking

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.

Prices · 18:30 UTC, 22 Sept
4 configurations in stock · 6 out of stock
MachineSpeedPer hourPer M tokensWhereRent
8× RTX PRO 6000768 GBcheapest~427 tok/sFaster than you read$16.54$10.75Vast.aiMarketplace · 98.4% reliableRent on Vast.ai
8× H100 PCIe640 GB~476 tok/sFaster than you read$19.16$11.16Vast.aiMarketplace · 98.3% reliableRent on Vast.ai
4× B200720 GB~1021 tok/sFaster than you read$25.00$6.80Vast.aiMarketplace · 99.6% reliableRent on Vast.ai
8× H2001128 GBrecommendedbest valuefastest~1144 tok/sFaster than you read$26.52$6.44Vast.aiMarketplace · 99.3% reliableRent on Vast.ai
No GPU at all

Use Kimi-K2-Thinking by the token

For one person chatting, that is about 2.6× cheaper than the best-value rented card above ($6.44 per million tokens). A rented GPU pays off when you keep it busy — many requests batched together, long agent runs — or when the data must stay on a machine you control, or the exact file you want is not served anywhere.

First time renting a GPU? How it works, in four steps
  1. Create an account and add credit. Both providers are prepaid: $10 is enough to try any card on this page for hours. A new RunPod account that arrives through a referral link gets a one-time $5 credit when it first adds $10.
  2. Pick the card and the count shown here, and a template: llama.cpp or vLLM for an OpenAI-compatible API, or one with a web chat if you only want to talk to the model. “How to launch it” below gives the exact image and command.
  3. Wait for the download. The weights are fetched on the machine; tens of gigabytes take a few minutes on a data-centre connection. You pay by the second from the moment it starts.
  4. Stop it when you are done. A running machine bills even when idle. A stopped one costs nothing per hour, though its disk is usually still billed until you delete it.
How to launch it

No Q4_K_M file of Kimi-K2-Thinking is catalogued here, so there is no exact command to vouch for. Start the machine from the llama.cpp server image ghcr.io/ggml-org/llama.cpp:server-cuda and point -hf at a Q4_K_M build from the publisher or a quantizer you trust — search Hugging Face for one. Check its file size against the memory figure on this page before you rent.

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-K2-Thinking three ways →

Under the hood

Architecture, read from the publisher’s file

The numbers every figure above is computed from, with the file they came from.

Architecture

✓ Architecture read from the published config.json

Retrieved 2026-09-01 at pinned commit a51ccc050d73; the native packed/fused checkpoint metadata is not used as a logical parameter total.

Architecture
kimi_k2
Layers
61
Hidden size
7,168
Attention heads
64
KV heads
64
Feed-forward width
18,432
Vocabulary
163,840
Context ceiling
262,144
RoPE theta
50,000
Experts
384
Experts / token
8
Expert width
2,048
Shared experts
1
Latent KV rank
512

Where the memory goes

tokenembedding× 61 decoder blocksLatent attentioncompressed KV rankfull contextRouted experts8 of 384 per tokenall residentoutputprojectiongrows with contextcapacity ≠ traffic

Latent attention projects keys and values into a compressed rank before caching them, which is why its KV cache is a fraction of a comparable dense model.

Open the exact official config.json ↗

More from moonshotai

Kimi-K2.5Kimi-K2.6Kimi-K2.7-CodeKimi-K2-InstructKimi-K2-Instruct-0905Kimi-K3Kimi-Linear-48B-A3B-Instruct

Use this answer in your own product

As JSON, from the API
curl -s https://llmbottleneck.com/v1/analyze \
  -H "Authorization: Bearer $LLMB_KEY" -H "content-type: application/json" \
  -d '{"model":"moonshotai-kimi-k2-thinking","quantization":"Q4_K_M","context":8192,"hardware":"nvidia-geforce-rtx-4090"}'

Same engine, same evidence, every field sourced. Free key in one step →

On your page, as a widget
<script src="https://llmbottleneck.com/widget.js"
  data-model="moonshotai-kimi-k2-thinking" 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 →

Citing this page

LLM Bottleneck. “Kimi-K2-Thinking VRAM and hardware requirements.” Architecture from moonshotai/Kimi-K2-Thinking at revision a51ccc050d73, retrieved 2026-09-01. https://llmbottleneck.com/models/moonshotai-kimi-k2-thinking

Every figure above is either the published value or a reconstruction whose measured error is on the accuracy page.

llmbottleneck
catalogue 2026-10-03models 327devices 135