llmrun
Moonshot AIKimi K3KimiK31ForConditionalGeneration

Kimi K3.1 Pro — Hardware Requirements & GPU Compatibility

VisionChat

Kimi K3.1 Pro is Moonshot AI's open-weight, natively multimodal agentic model, post-trained from Kimi K3 with large-scale reinforcement learning and aimed at coding, financial research and other agent workloads. It is a Mixture-of-Experts model with 2.8 trillion parameters in total and 104 billion activated per token, built from 93 layers with 896 experts (16 selected per token, plus 2 shared). It mixes Kimi Delta Attention layers with Gated DeepSeek Sparse Attention layers for more efficient long-context modeling, and pairs the language model with a 401M-parameter MoonViT-V2 vision encoder. Context length is 1,048,576 tokens. It is released under the custom Kimi K3.1 Pro License, so terms are not a standard open-source license and should be read before commercial use. The weights are trained with MXFP4 quantization-aware training, but at this scale it is server-class hardware: expect a multi-GPU node, not a consumer machine.

2780.1B params · MoE106B active1049K contextOther41.0K downloads2.0K likes6.3K quant downloads

Kimi K3.1 Pro runs on a GPU cluster

Q4_K_M needs ~1.7 TB · about 3 nodes of 8× H100 SXM 80 GB

Score 139.8#1 of 107 open#1 of 107 that fit ≤2 TBOther

Not set up for it?

Reference GPU · RTX 4090 24 GB

Estimates, ±20% · how we estimate

SOn your NVIDIA GB200 NVL72: Q4_K_M fits · 11679.9 GB free · ~170 tok/s · up to 1M contextNVIDIA GB200 NVL72 page →

llmrun Composite Score →

139.8

Benchmarks · tier chart ↓
Among 107 open modelsHow the Composite Score is built

Run Kimi K3.1 Pro

Best fit for your NVIDIA GB200 NVL72. Every other method, OS and quant is in the step-by-step guide.Each way opens a step-by-step guide for your OS and quant.

Recommended for your NVIDIA GB200 NVL72

Ollama
one command · needs Ollama installed
$ollama run hf.co/KarLexAI/Kimi-K3.1-Pro-GGUF

Q4_K_M fits fully · 11679.9 GB free · 16K context

Download: Kimi-K3.1-Pro-Q4_K_M.gguf1668.07 GBKarLexAI

install, OS, other quants

Ways to runOther ways

each opens step-by-step

Run in cloud

Rent 3× 8-GPU H100 SXM nodes · Q4_K_M

Estimated hourly price
~$64.56/hr
est., whole cluster
GPUs
24 GPUs
1.9 TB of VRAM
Deployon RunPodRent H100 SXM$2.69/hr per GPU (opens in new tab)

$5+ bonus for new users

Compare GPUs →

How Much VRAM Does Kimi K3.1 Pro Need?

Select a quantization level; the listed hardware adheres to it.

QuantizationVRAM+32KOn your device
770.3 GB852.2 GB2-bit · i-quant extreme compression — Importance-weighted 2-bit, extreme compression — significant quality lossFits: 1M context · ~172 tok/s
1187.3 GB1269.2 GB2-bit · K-quant — 2-bit quantization with K-quant improvementsFits: 1M context · ~161 tok/s
1361.1 GB1443 GB3-bit medium — 3-bit medium quantizationFits: 1M context · ~157 tok/s
1673.8 GB1755.7 GB4-bit medium — 4-bit medium quantization — most popular sweet spotFits: 1M context · ~170 tok/s
1986.6 GB2068.5 GB5-bit medium — 5-bit medium quantization — good quality/size tradeoffFits: 1M context · ~166 tok/s
2299.3 GB2381.3 GB6-bit, very good — 6-bit quantization, very good qualityFits: 1M context · ~162 tok/s
2785.9 GB2867.8 GB8-bit, near-lossless — 8-bit quantization, near-losslessFits: 1M context · ~157 tok/s
5566.0 GB5647.9 GB16-bit · bfloat — Brain floating point 16 — preferred for trainingFits: 1M context · ~157 tok/s
est.calculated, no published file yet

Selected Q4_K_M

4-bit medium quantization

1712.1GBat 16K context

Weights
1668.1 GB
KV cache
43.7 GB
Overhead
0.3 GB

Fits NVIDIA GB200 NVL72 · 11679.9 GB free · ~170 tok/s

Which Hardware Can Run Kimi K3.1 Pro?

Q4_K_M · 1673.8 GB

Kimi K3.1 Pro (Q4_K_M) requires 1673.8 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 2176+ GB is recommended. Using the full 1049K context window can add up to 2790.6 GB, bringing total usage to 4464.4 GB. No consumer GPU has enough memory.

Needs server-class hardware

Kimi K3.1 Pro runs on a GPU cluster · Q4_K_M needs ~1.7 TB

Hardware that runs this model
HardwareQuantMemoryContextHardware page
8× H200 SXM 141 GB · 1.1 TBIQ2_XXS808.5 GB64K ctxH200 SXM →IQ2_XXS · 808.5 GB · 64K ctx
8× B200 192 GB · 1.5 TBQ2_K1225.5 GB64K ctxB200 →Q2_K · 1225.5 GB · 64K ctx
GPU cluster · 3 nodes of 8× H100 SXM 80 GBQ4_K_M1712.1 GB16K ctxH100 SXM →Q4_K_M · 1712.1 GB · 16K ctx

Also: AMD Instinct MI355X Platform (8x MI355X) · NVIDIA DGX B300 · NVIDIA GB200 NVL72 · NVIDIA GB300 NVL72

≤ 20736 GB tier →

By VRAM tier · typical GPUs and Macs at Q4_K_M

Consumer GPUs (up to 32 GB) and Macs up to 512 GB: too small for this model (needs ≥ 808.5 GB)8 GB tier →12 GB tier →16 GB tier →24 GB tier →Macs →

How good is Kimi K3.1 Pro?

The llmrun Score merges 10 public benchmark boards into a single figure, where 100 represents the average for open models.

Score vs VRAM in the ≤2 TB tier

  • Kimi K3.1 Pro
  • Other models
  • Best Score for its VRAM

#1 of 107 that fit ≤2 TBEvery model here fits your NVIDIA GB200 NVL72 (13392 GB)

Each dot is a model at the best quant that fits ≤2 TB. The line steps up to the best Score available at each amount of VRAM; a dot below it is beaten by a model that needs no more memory.

Full ≤2 TB ranking →

Benchmarks

The 6 boards that weigh most in the llmrun Score, of 17 with a result for this model; badges show who ran the test.

  • LMArena Text

    LMArena · Nov 2026Arena

    1484.7

    #2/191 open

    Arena
  • LMArena WebDev

    LMArena · Nov 2026Arena

    1687.2

    #1/58 open

    Arena
  • SciCode

    Epoch AI · Oct 2026Lab-reported

    60.8%

    #2/74 open

    Lab-reported
  • WeirdML v3

    Epoch AI · Oct 2026Lab-reported

    9.9%

    #1/6 open

    Lab-reported
  • ALE-Bench

    Epoch AI · Oct 2026Lab-reported

    1932.8

    #1/51 open

    Lab-reported
  • LiveBench 2026 Coding

    LiveBench · Oct 2026Independent

    81.8

    #8/20 open

    Independent
All 17 Kimi K3.1 Pro benchmarks

In the ≤2 TB class

107 models fit ≤2 TB

Same Score, only models that fit ≤2 TB at their best quant: this model's peers by memory need.

  1. #1

    Kimi K3.1 ProMoonshot AI

    139.8

    1712.1 GB

  2. #2

    Inkling 2 SmallThinking Machines

    134.7

    359.1 GB

  3. #3

    DeepSeek V4.1 Flash 1017DeepSeek

    134.5

    458.4 GB

  4. #4

    MiMo V2.6 Pro RLXiaomi

    133.5

    614.1 GB

  5. #5

    Kimi K3Moonshot AI

    132.8

    1670.0 GB

  6. #6

    MiMo V2.6 Flash RLXiaomi

    131.9

    186.5 GB

Full ≤2 TB ranking →

Alternatives to Kimi K3.1 Pro

Kimi K3 family · 3 models →

Official models with an llmrun Score, each answering one question; sizes are each model's default quant (usually Q4_K_M) with a 16K-token context.

Similar models

4
ModelScoreVRAMWhy it is similar
DeepSeek V4.1 Flash 1017DeepSeek134.5458.4 GBsimilar score
MiMo V2.6 Pro RLXiaomi133.5614.1 GBsimilar score
MiMo V2.6 Flash RLXiaomi131.9186.5 GBsimilar score
DeepSeek V4 ProDeepSeek116.7959.8 GBsmaller
Fine-tunes & variants(2)

Specifications

Publisher
Moonshot AI
Family
Kimi K3
Parameters
2780.1B
Active parameters
106B per token
Architecture
KimiK31ForConditionalGeneration
Context Length
1,048,576 tokens
Vocabulary Size
163,840
Release Date
2026-11-02
License
Other

Resources

Frequently Asked Questions

How much VRAM does Kimi K3.1 Pro need?

Kimi K3.1 Pro requires 1673.8 GB of VRAM at Q4_K_M, or 5566.0 GB at BF16. Full 1049K context adds up to 2790.6 GB (4464.4 GB total).

VRAM = Weights + KV Cache + Overhead

Weights = 2780.1B × 4.8 bits ÷ 8 = 1668.1 GB

KV Cache + Overhead ≈ 5.7 GB (at 2K context + ~0.3 GB framework)

Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.

KV Cache + Overhead ≈ 2796.3 GB (at full 1049K context)

VRAM usage by quantization

1673.8 GB
4464.4 GB

Learn more about VRAM estimation →

Can NVIDIA GeForce RTX 5090 run Kimi K3.1 Pro?

No — Kimi K3.1 Pro requires at least 770.3 GB at IQ2_XXS, which exceeds the NVIDIA GeForce RTX 5090's 32 GB of VRAM.

What's the best quantization for Kimi K3.1 Pro?

For Kimi K3.1 Pro, Q4_K_M (1673.8 GB) offers the best balance of quality and VRAM usage. Q5_K_M (1986.6 GB) provides better quality if you have the VRAM. The smallest option is IQ2_XXS at 770.3 GB.

VRAM requirement by quantization

IQ2_XXS
770.3 GB
Q3_K_M
1361.1 GB
Q4_K_M ★
1673.8 GB
Q5_K_M
1986.6 GB
Q6_K
2299.3 GB
BF16
5566.0 GB

★ Recommended — best balance of quality and VRAM usage.

Learn more about quantization →

Can I run Kimi K3.1 Pro on a Mac?

Kimi K3.1 Pro requires at least 770.3 GB at IQ2_XXS. None of the 159 Macs we list has that much usable memory (the largest has 508.5 GB after macOS overhead).

Can I run Kimi K3.1 Pro locally?

Yes — Kimi K3.1 Pro can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.

What's the download size of Kimi K3.1 Pro?

At Q4_K_M, the download is about 1668.07 GB. The full-precision BF16 version is 5560.22 GB. The smallest option (IQ2_XXS) is 764.53 GB.

Which GPUs can run Kimi K3.1 Pro?

No single consumer GPU has enough VRAM to run Kimi K3.1 Pro at Q4_K_M (1673.8 GB). Multi-GPU or professional hardware is required.

Which devices can run Kimi K3.1 Pro?

Kimi K3.1 Pro requires at least 1673.8 GB at Q4_K_M, which exceeds the usable memory of every unified-memory device we list (the largest has 508.5 GB).