Skip to content
localllm
日本語

moonshotai

Kimi-Dev-72B

Parameters
72.7B
Max context length
131,072
Architecture
qwen2
License
Commercial OK
Base model
moonshotai/Kimi-Dev-72B
License
mit

About Kimi-Dev-72B

Kimi-Dev-72B is a GGUF model based on moonshotai/Kimi-Dev-72B on Hugging Face. GGUF files are published by unsloth/Kimi-Dev-72B-GGUF. It has about 72.7B parameters, the qwen2 architecture, a maximum context length of 131,072 tokens (values written in the GGUF).

There are 19 quantizations, with file sizes from 23.0 GB (UD-IQ1_S) to 145 GB (BF16) (actual sizes on Hugging Face).

Running UD-IQ1_S (23.0 GB) with an 8,192-token context needs an estimated 26.6 GB of memory, and an estimated 34.6 GB at 32,768 tokens. For 8,192 tokens, a GPU with 32 GB or more of VRAM (such as GeForce RTX 5090), or a Mac with 32 GB or more of unified memory, is a rough guide (macOS caps how much memory the GPU can use, so leave some headroom).

The original model is licensed under MIT License, which allows commercial use (keep the license text and copyright notice when redistributing). Check the original license for the terms of use.

File size and estimated memory per quantization

  • Up to 8 GB
  • Up to 16 GB
  • Up to 24 GB
  • Up to 48 GB
  • Up to 128 GB
  • More

File sizes are the actual values on Hugging Face. Memory is estimated for each context length (KV cache f16). The formula is in How memory is calculated.

All memory figures are Estimated (KV cache f16). It should run on a setup with at least this much VRAM (or Mac unified memory). Click a value to see the devices and models that fit that amount (Find by GPU / Mac).

About the "quality guide" (how far a quantization drifts from the original)

For each quantization type, how far the output drifts from the original model, grouped into six levels from values llama.cpp measured on Llama-3-8B.

The value is mean KLD (the difference between the original and quantized model in next-token probabilities); smaller means closer to the original. Quantizations made with an imatrix (importance data) come out slightly smaller. Levels use the value without an imatrix (or with one, for types that only have that).

These values were not measured on this model. Larger models and MoE models tend to drift less with the same type. Distributor-specific mixed types (such as UD-Q4_K_XL) cannot be matched by type name alone, so they show "-".

Quantization Quality guide KLD (no imatrix) KLD (imatrix) Size (Llama-3-8B)
Q8_0 Nearly identical 0.0014 - 7.96 GiB
Q6_K Very small 0.0055 - 6.14 GiB
Q5_K_M Very small 0.0108 - 5.33 GiB
Q5_K_S Small 0.0166 - 5.21 GiB
Q5_1 Small 0.0180 - 5.65 GiB
Q5_0 Small 0.0222 - 5.21 GiB
Q4_K_M Small 0.0313 0.0282 4.58 GiB
Q4_K_S Medium 0.0431 0.0320 4.37 GiB
IQ4_NL Small - 0.0357 4.35 GiB
IQ4_XS Small - 0.0363 4.14 GiB
Q4_1 Medium 0.0717 - 4.78 GiB
Q4_0 Medium 0.0719 - 4.34 GiB
Q3_K_L Medium 0.0872 0.0731 4.03 GiB
Q3_K_M Medium 0.1019 0.0844 3.74 GiB
IQ3_M Medium - 0.1025 3.53 GiB
IQ3_S Medium - 0.1113 3.42 GiB
IQ3_XS Large - 0.1387 3.28 GiB
IQ3_XXS Large - 0.1836 3.05 GiB
Q3_K_S Large 0.2282 0.1998 3.41 GiB
IQ2_M Large - 0.3260 2.74 GiB
Q2_K Very large 0.4451 0.3322 2.96 GiB
Q2_K_S Very large - 0.4034 2.96 GiB
IQ2_S Very large - 0.4392 2.56 GiB
IQ2_XS Very large - 0.5463 2.43 GiB
IQ2_XXS Very large - 0.8120 2.24 GiB
IQ1_M Very large - 1.3931 2.01 GiB
IQ1_S Very large - 2.2546 1.88 GiB

Source: llama.cpp (MIT), measured on Llama-3-8B against FP16, Wikitext-2, revision f364eb6f. Original table · Reading and choosing quantizations

GGUF source

unsloth/Kimi-Dev-72B-GGUF

Converted and published by a third party About this label

Last checked: Oct 2, 2026

Architecture
qwen2
Supported before llama.cpp b4409
Parameters
72.7B
Max context length
131,072
License
mit Commercial OK
File size and memory per quantization for Kimi-Dev-72B (unsloth/Kimi-Dev-72B-GGUF)
Quantization File size Measured Quality guide 4,096 8,192 16,384 32,768 65,536 131,072 max
UD-IQ1_S 23.0 GB - 25.2 GB 26.6 GB 29.3 GB 34.6 GB 45.4 GB 66.8 GB
UD-IQ1_M 24.0 GB - 26.2 GB 27.5 GB 30.2 GB 35.6 GB 46.3 GB 67.8 GB
UD-IQ2_XXS 25.7 GB - 27.9 GB 29.2 GB 31.9 GB 37.3 GB 48.0 GB 69.5 GB
UD-IQ2_M 29.6 GB - 31.8 GB 33.1 GB 35.8 GB 41.2 GB 51.9 GB 73.4 GB
UD-Q2_K_XL 30.3 GB - 32.5 GB 33.9 GB 36.6 GB 41.9 GB 52.7 GB 74.1 GB
UD-IQ3_XXS 31.9 GB - 34.0 GB 35.4 GB 38.1 GB 43.4 GB 54.2 GB 75.7 GB
Q3_K_S 34.5 GB Large 36.7 GB 38.0 GB 40.7 GB 46.1 GB 56.8 GB 78.3 GB
Q3_K_M 37.7 GB Medium 39.9 GB 41.2 GB 43.9 GB 49.3 GB 60.0 GB 81.5 GB
UD-Q3_K_XL 38.3 GB - 40.5 GB 41.8 GB 44.5 GB 49.9 GB 60.6 GB 82.1 GB
IQ4_XS 39.7 GB Small 41.9 GB 43.3 GB 46.0 GB 51.3 GB 62.1 GB 83.5 GB
IQ4_NL 41.3 GB Small 43.5 GB 44.8 GB 47.5 GB 52.9 GB 63.6 GB 85.1 GB
Q4_0 41.4 GB Medium 43.6 GB 44.9 GB 47.6 GB 53.0 GB 63.7 GB 85.2 GB
Q4_1 45.7 GB Medium 47.9 GB 49.2 GB 51.9 GB 57.3 GB 68.0 GB 89.5 GB
UD-Q5_K_XL (2 parts) 54.1 GB - 56.3 GB 57.7 GB 60.4 GB 65.7 GB 76.5 GB 97.9 GB
Q6_K (2 parts) 64.3 GB Very small 66.5 GB 67.9 GB 70.6 GB 75.9 GB 86.7 GB 108 GB
UD-Q6_K_XL (2 parts) 67.2 GB - 69.4 GB 70.7 GB 73.4 GB 78.7 GB 89.5 GB 111 GB
Q8_0 (2 parts) 77.3 GB Nearly identical 79.5 GB 80.8 GB 83.5 GB 88.8 GB 99.6 GB 121 GB
UD-Q8_K_XL (2 parts) 84.0 GB - 86.2 GB 87.5 GB 90.2 GB 95.6 GB 106 GB 128 GB
BF16 (3 parts) 145 GB Original precision 148 GB 149 GB 152 GB 157 GB 168 GB 189 GB 
Ollama  ollama run hf.co/unsloth/Kimi-Dev-72B-GGUF:UD-IQ1_S
llama.cpp  llama-server -hf unsloth/Kimi-Dev-72B-GGUF:UD-IQ1_S

Generation speed on common GPUs and Macs (estimated)

A rough guide to how fast UD-IQ1_S (23.0 GB) generates text. Each token reads the model weights once, so the upper limit is memory bandwidth ÷ file size. The guide is 60–80% of that, where most published measurements fall.

Device Memory bandwidth Guide (tokens/s) Estimated Upper limit
GeForce RTX 5090 32 GB 1,792 GB/s 47–62 78
Apple M3 Pro 36 GB 150 GB/s 3.9–5.2 6.5
Apple M1 Max 64 GB 400 GB/s 10–14 17
Apple M5 Ultra 96 GB 1,200 GB/s 31–42 52

Estimates only. Prompt processing speed (limited by compute) is not included. Longer contexts and offloading part of the model to the CPU make it slower. Macs are shown only when installed memory is at least 4/3 of what is needed. Devices without published bandwidth (some GeForce cards, etc.) are not shown.

License and commercial use

MIT License Commercial OK

Allows commercial use, modification and redistribution. When redistributing, keep the copyright notice and license text.

Based on the license on the original model's Hugging Face page. Original license text (checked Oct 18, 2026)

The summary states facts this site read from the original text and is not legal advice. Always check the original license and its latest terms before use. All licenses

How to run

Example commands using UD-IQ1_S. To use another quantization, replace the name with one from the table above. How to run local LLMs

Ollama

Run it by pointing directly at the GGUF on Hugging Face.

ollama run hf.co/unsloth/Kimi-Dev-72B-GGUF:UD-IQ1_S

llama.cpp

Pass the repository and quantization name to llama-server (or llama-cli) with -hf.

llama-server -hf unsloth/Kimi-Dev-72B-GGUF:UD-IQ1_S

LM Studio

Paste unsloth/Kimi-Dev-72B-GGUF into the search box, download it, and choose the quantization from the file list.

Frequently asked questions

How much memory do I need to run Kimi-Dev-72B (UD-IQ1_S)?
The file size is 23.0 GB (measured). With a 8,192-token context, the estimated memory including the KV cache is 26.6 GB. The formula is published at https://localllm.az-arte.com/en/guide/memory-calculation.
Can Kimi-Dev-72B be used commercially?
The original model's license is MIT License (Commercial use allowed).Allows commercial use, modification and redistribution. When redistributing, keep the copyright notice and license text. Check the original license text before deciding.
Does Kimi-Dev-72B run in 8 GB, 12 GB, 16 GB or 24 GB of VRAM?
8 GB: no quantization fits; 12 GB: no quantization fits; 16 GB: no quantization fits; 24 GB: no quantization fits (estimated at 8,192 tokens with an f16 KV cache, files from the most-downloaded source).
What is the command to run Kimi-Dev-72B with Ollama?
Run "ollama run hf.co/unsloth/Kimi-Dev-72B-GGUF:UD-IQ1_S" to use the GGUF on Hugging Face directly. To use another quantization, replace the name at the end.