Moonshot/Mixture of Experts

MKimi-Linear-48B-A3B

chatcodingreasoningmultilingual
48B
Parameters (3B active)
1024K
Context length
1
Benchmarks
14
Quantizations
48K
HF downloads
Architecture
MoE
Released
2025-10-30
Layers
27
KV Heads
32
Head Dim
72
Family
kimi

Quantization Options

QuantBitsVRAMQuality
IQ3_XXS3.2520.0 GBlow
IQ3_XS3.521.5 GBlow
Q3_K_S3.6422.3 GBlow
IQ3_M3.7623.0 GBlow
Q3_K_M424.5 GBlow
Q3_K_L4.326.3 GBmoderate
IQ4_XS4.4627.2 GBmoderate
Q4_K_S4.6728.5 GBmoderate
Q4_K_M4.8929.8 GBgood
Q5_K_S5.5733.9 GBgood
Q5_K_M5.734.7 GBgood
Q6_K6.5639.8 GBexcellent
Q8_08.551.5 GBlossless
FP161696.5 GBlossless

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

Benchmarks (1)

MMLU-PRO51.0

Run this model

Easiest way to get starteddocs →
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

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
Apple M1 Max (32GB)
32 GB VRAM • 400 GB/s
APPLE
$1499
Apple M2 Max (32GB)
32 GB VRAM • 400 GB/s
APPLE
$1799
NVIDIA V100 SXM2 32GB
32 GB VRAM • 900 GB/s
NVIDIA
$3500
Apple M2 Pro (32GB)
32 GB VRAM • 200 GB/s
APPLE
$1499
Apple M4 (32GB)
32 GB VRAM • 120 GB/s
APPLE
$1199
NVIDIA Tesla V100 DGXS 32 GB
32 GB VRAM • 897 GB/s
NVIDIA
NVIDIA Tesla V100 PCIe 32 GB
32 GB VRAM • 897 GB/s
NVIDIA
NVIDIA Tesla V100 SXM2 32 GB
32 GB VRAM • 898 GB/s
NVIDIA
NVIDIA Tesla V100 SXM3 32 GB
32 GB VRAM • 981 GB/s
NVIDIA
AMD Radeon Instinct MI60
32 GB VRAM • 1020 GB/s
AMD
NVIDIA Tesla V100S PCIe 32 GB
32 GB VRAM • 1130 GB/s
NVIDIA
AMD Radeon Instinct MI100
32 GB VRAM • 1230 GB/s
AMD
$5000
NVIDIA RTX 5000 Ada Generation
32 GB VRAM • 576 GB/s
NVIDIA
$4000
NVIDIA GeForce RTX 5090
32 GB VRAM • 1790 GB/s
NVIDIA
$1999
NVIDIA GeForce RTX 5090 D
32 GB VRAM • 1790 GB/s
NVIDIA
$1999
NVIDIA Jetson AGX Xavier 32 GB
32 GB VRAM • 136 GB/s
NVIDIA
NVIDIA Quadro GV100
32 GB VRAM • 868 GB/s
NVIDIA
NVIDIA TITAN V CEO Edition
32 GB VRAM • 868 GB/s
NVIDIA
NVIDIA Tesla PG500-216
32 GB VRAM • 1130 GB/s
NVIDIA
NVIDIA Tesla PG503-216
32 GB VRAM • 1130 GB/s
NVIDIA
AMD Radeon Pro Vega II
32 GB VRAM • 825 GB/s
AMD
AMD Radeon Pro Vega II Duo
32 GB VRAM • 1020 GB/s
AMD
AMD Radeon PRO V620
32 GB VRAM • 512 GB/s
AMD
AMD Radeon PRO W6800
32 GB VRAM • 512 GB/s
AMD
$2249
AMD Radeon Pro W6800X
32 GB VRAM • 512 GB/s
AMD
AMD Radeon Pro W6800X Duo
32 GB VRAM • 512 GB/s
AMD
AMD Radeon Pro W6900X
32 GB VRAM • 512 GB/s
AMD
NVIDIA Jetson AGX Orin 32 GB
32 GB VRAM • 205 GB/s
NVIDIA
AMD Radeon PRO W7800
32 GB VRAM • 576 GB/s
AMD
$2499

Find the best GPU for Kimi-Linear-48B-A3B

Build Hardware for Kimi-Linear-48B-A3B

Kimi-Linear-48B-A3B48B Parameter Mixture of Experts LLM

Model Specifications

Parameters
48B (3B active)
Architecture
Mixture of Experts
Context Length
1024K tokens
Capabilities
chat, coding, reasoning, multilingual
Release Date
2025-10-30
Provider
Moonshot
Family
kimi

VRAM Requirements

QuantizationBPWVRAMQuality
IQ3_XXS3.2520.0 GB82%
IQ3_XS3.521.5 GB84%
Q3_K_S3.6422.3 GB85%
IQ3_M3.7623.0 GB86%
Q3_K_M424.5 GB88%
Q3_K_L4.326.3 GB90%
IQ4_XS4.4627.2 GB92%
Q4_K_S4.6728.5 GB93%
Q4_K_M4.8929.8 GB94%
Q5_K_S5.5733.9 GB96%
Q5_K_M5.734.7 GB96%
Q6_K6.5639.8 GB97%
Q8_08.551.5 GB100%
FP161696.5 GB100%

Benchmark Scores

MMLU-PRO51.0

How to Run Kimi-Linear-48B-A3B

Run Kimi-Linear-48B-A3B locally with Ollama (needs 29.8 GB VRAM at Q4_K_M):

ollama run kimi-linear:48b-a3b

Compatible GPUs (30)

GPUs that can run Kimi-Linear-48B-A3B at Q4_K_M quantization: