Moonshot/Mixture of Experts

MKimi-Linear-48B-A3B

Name: Kimi-Linear-48B-A3B
Author: Moonshot

Kimi Linear is a hybrid linear attention architecture that outperforms traditional full attention methods across various contexts, including short, long, and reinforcement learning (RL) scaling regimes.

chatcodingreasoningmultilingual

48B

Parameters (3B active)

1024K

Context length

Benchmarks

Quantizations

48K

HF downloads

Architecture

MoE

Released

2025-10-30

Layers

KV Heads

Head Dim

Family

kimi

Quantization Options

Quant	Bits	VRAM	Quality
IQ3_XXS	3.25	20.0 GB	low
IQ3_XS	3.5	21.5 GB	low
Q3_K_S	3.64	22.3 GB	low
IQ3_M	3.76	23.0 GB	low
Q3_K_M	4	24.5 GB	low
Q3_K_L	4.3	26.3 GB	moderate
IQ4_XS	4.46	27.2 GB	moderate
Q4_K_S	4.67	28.5 GB	moderate
Q4_K_M	4.89	29.8 GB	good
Q5_K_S	5.57	33.9 GB	good
Q5_K_M	5.7	34.7 GB	good
Q6_K	6.56	39.8 GB	excellent
Q8_0	8.5	51.5 GB	lossless
FP16	16	96.5 GB	lossless

Select your GPU above to see speed estimates and compatibility for each quantization.

▸ READY TO RUN THIS?RENT BY THE HOUR

RENT A GPU AND RUN KIMI-LINEAR-48B-A3B NOW

Rent on RunPod →Or Vast.ai →

Spin up an A100 / H100 / 4090 in ~60s. Pay by the second. Cancel anytime.

Community Ratings

Loading ratings...

Benchmarks (12)

MMLU-PRO51.0

GPQA Diamond41.2

LiveCodeBench37.8

AIME36.3

AA Math36.3

IFBench28.1

AA Long Context25.7

SciCode19.9

AA Intelligence14.4

AA Coding14.2

Terminal-Bench11.4

HLE2.7

Run this model

▸Easiest way to get started·Beginners

DOCS ↗

curl -fsSL https://ollama.com/install.sh | sh

$ollama run kimi-linear:48b-a3b-q4_K_M

Downloads and runs automatically. Add --verbose for speed stats.

▸ SETUP GUIDE

Auto-setup with fitmyllm CLI

Detects your GPU, recommends the best model, downloads it, and starts chatting — zero config. Benchmarks your speed and contributes anonymous data to improve predictions.

pip install fitmyllmthen run fitmyllmLearn more

Auto-detect GPULive tok/s in chatSpeed benchmarks9 inference engines

HuggingFace Ollama Library GGUF Downloads Build Hardware

GPUs that can run this model

At Q4_K_M quantization. Sorted by minimum VRAM.

NVIDIA RTX 5090

32 GB VRAM • 1792 GB/s

NVIDIA

$1999