Updated minutes ago
ReleasedJuly 16, 2026Verified — · huggingface.co
Moonshot

Kimi K3

Moonshot AI · moe · 2779.9B parameters · 1,048,576 context

Quality
50.0
Deploy this model →Use via API →Calculate ROI →Compare with othersFine-tune this model

Parameters

2.8T

Context Window

1024K tokens

Architecture

MoE

Best GPU

B200 SXM

Cheapest API

$15.00/M

Intelligence Brief

Kimi K3 is a 2779.9B parameter Mixture-of-Experts (896 experts, 16 active) model from Moonshot AI, featuring Multi-Head Attention (MHA) with 93 layers and 7,168 hidden dimensions. With a 1,048,576 token context window, it supports tools, vision, structured output, code, math, multilingual, reasoning. The most cost-effective API deployment is via openrouter at $15.00/M output tokens. For self-hosted inference, B200 SXM delivers optimal throughput at $136352/month.

Provider pricing

1 provider · canonical: openrouter
Provider Input $/M Output $/M ▲Notes
openroutercanonical$3.00$15.00cheapest input · cheapest output

Prices update via the nightly pricing cron + admin approvals at /admin/ingest-queue. The leaderboard's Input/Output cells show the canonical rate above; this table shows the full spread.

Recent changes

Loading…

Related models

5 suggestions

Picks: same family first, then same vendor within ±2× params, then top tag-overlap matches. Price shown is the cheapest Output $/M across providers — the row's page shows the canonical anchor.

Architecture Details

TypeMOE
Total Parameters2779.9B
Active Parameters32B
Layers93
Hidden Dimension7,168
Attention Heads96
KV Heads96
Head Dimension75
Vocab Size163,840
Total Experts896
Active Experts16

Memory Requirements

BF16 Weights

5178.0 GB

FP8 Weights

2589.0 GB

INT4 Weights

1294.5 GB

KV-Cache per Token2678400 bytes
Activation Estimate0.00 GB

Fits on (multi-GPU with Tensor Parallelism)

RECOMMENDEDx6B300288GB eachINT4x7Instinct MI325X256GB eachINT4x8Groq LPU230GB eachINT4x9GB200 NVL72 (per GPU)192GB eachINT4x9GB300 NVL72 (per GPU)192GB eachINT4x52TPU v432GB eachINT4x52TPU v6e (Trillium)32GB eachINT4x103TPU v5e16GB eachINT4

Multi-GPU configurations use Tensor Parallelism (TP) to split model layers across GPUs. Requires NVLink or NVSwitch interconnect for optimal performance.

This model requires multi-GPU deployment. Minimum: 6x B300 (288GB each) with Tensor Parallelism.

GPU Compatibility Matrix

Kimi K3 is compatible with 0% of GPU configurations across 41 GPUs at 3 precision levels.

BF16 (Full)
FP8 (Half)
INT4 (Quarter)
Blackwell(7 GPUs)
B200 NVL (pair)360GB
✗
✗
✗
B300288GB
✗
✗
✗
B100 SXM192GB
✗
✗
✗
GB200 NVL72 (per GPU)192GB
✗
✗
✗
Hopper(7 GPUs)
H100 NVL 94GB (per GPU pair)188GB
✗
✗
✗
H200 SXM141GB
✗
✗
✗
H2096GB
✗
✗
✗
GH20096GB
✗
✗
✗
Ada Lovelace(11 GPUs)
L40S48GB
✗
✗
✗
L4048GB
✗
✗
✗
RTX 6000 Ada48GB
✗
✗
✗
L2048GB
✗
✗
✗
Ampere(16 GPUs)
A100 80GB SXM80GB
✗
✗
✗
A100 80GB PCIe80GB
✗
✗
✗
A1664GB
✗
✗
✗
RTX A600048GB
✗
✗
✗
Legend:No fitVery tightTightModerateGoodExcellent

GPU Recommendations

B200 SXMoptimal

FP8 · 32 GPUs · tensorrt-llm

83/100

score

Throughput

140.0 tok/s

Latency (ITL)

7.1ms

Est. TTFT

1ms

Cost/Month

$136352

Cost/M Tokens

$370.60

Use this config →
B200 NVL (pair)optimal

FP8 · 16 GPUs · tensorrt-llm

83/100

score

Throughput

140.0 tok/s

Latency (ITL)

7.1ms

Est. TTFT

1ms

Cost/Month

$159432

Cost/M Tokens

$433.33

Use this config →
H200 SXMoptimal

FP8 · 32 GPUs · tensorrt-llm

80/100

score

Throughput

140.0 tok/s

Latency (ITL)

7.1ms

Est. TTFT

1ms

Cost/Month

$81690

Cost/M Tokens

$222.03

Use this config →

Deployment Options

API

API Deployment

openrouter

$15.00/M

output tokens

Self-Hosted

Single GPU

Requires multi-GPU setup (2589 GB VRAM needed)

Scale

Multi-GPU

B200 SXM x32

140.0 tok/s

TP· $136352/mo

Open Full Calculator →

API Pricing Comparison

ProviderInput $/MOutput $/MBadges
openrouter$3.00$15.00
Cheapest

Cost Analysis

ProviderInput $/MOutput $/M~Monthly Cost
openrouterBest Value$3.00$15.00$90

Cost per 1,000 Requests

Short (500 tok)

$4.50

via openrouter

Medium (2K tok)

$18.00

via openrouter

Long (8K tok)

$54.00

via openrouter

Performance Estimates

Throughput by GPU

B200 SXM
140.0 tok/s
B200 NVL (pair)
140.0 tok/s
H200 SXM
140.0 tok/s

VRAM Breakdown (B200 SXM, FP8)

Weights
KV
Weights 86.9 GBKV-Cache 21.9 GBActivations 0.0 GBOverhead 4.3 GB

Precision Impact

bf16

173.7 GB

weights/GPU

fp8

86.9 GB

weights/GPU

~140.0 tok/s

int4

43.4 GB

weights/GPU

Capabilities

Features

✓ Tool Use✓ Vision✓ Code✓ Math✓ Reasoning✓ Multilingual✓ Structured Output

Supported Frameworks

vllmsglang

Supported Precisions

BF16 (default)FP8INT4

Where to Deploy Kimi K3

Similar Models

Kimi K2.7 Code

1026.9B params · moe

Quality: 50

from $3.50/M

Smaller context, Cheaper, Smaller modelCompare →

Kimi K2.5

1000B params · moe

Quality: 54

from $0.00/M

Smaller context, Cheaper, Smaller modelCompare →

Llama 4 Behemoth

2000B params · moe

Quality: 93

from $16.00/M

Higher qualityCompare →

GPT-4.5 Preview

1500B params · moe

Quality: 93

from $150.00/M

Smaller context, Higher quality, More expensiveCompare →

Frequently Asked Questions

How much VRAM does Kimi K3 need for inference?

Kimi K3 requires approximately 5178.0 GB of VRAM at BF16 precision, 2589.0 GB at FP8, or 1294.5 GB at INT4 quantization. Additional VRAM is needed for KV-cache (2678400 bytes per token) and activations (~0.00 GB).

What is the best GPU for Kimi K3?

The top recommended GPU for Kimi K3 is the B200 SXM (x32) using FP8 precision. It achieves approximately 140.0 tokens/sec at an estimated cost of $136352/month ($370.60/M tokens). Score: 83/100.

How much does Kimi K3 inference cost?

Kimi K3 API inference starts from $3.00/M input tokens and $15.00/M output tokens. Self-hosted inference costs depend on your GPU configuration — use our ROI calculator for a detailed breakdown.