What are the hardware requirements for Kimi K2 Thinking 3bit?

Kimi K2 Thinking 3bit requires approximately 219.3 GB of VRAM and supports a context window of 256K tokens. Quantized variants may run on less VRAM; see the Quantized Models section on this page.

Who developed Kimi K2 Thinking 3bit and how large is it?

Kimi K2 Thinking 3bit is developed by mlx-community, a model with 1026.4b parameters. The model is published as open weights on Hugging Face and indexed on LLM Explorer with full benchmark history.

How does Kimi K2 Thinking 3bit perform on standard benchmarks?

Kimi K2 Thinking 3bit has the following published scores: LMSYS ELO 1430. Compare against reference models on this page or on the LLM Explorer leaderboards.

Kimi K2 Thinking 3bit by mlx-community — VRAM 219.3GB, 256K context

Name: Kimi K2 Thinking 3bit
Author: mlx-community

Kimi K2 Thinking 3bit is an open-source language model by mlx-community. Features: 1026.4b LLM, VRAM: 219.3GB, Context: 256K, License: other, Quantized, LLM Explorer Score: 0.36, ELO: 1430.

3-bit 3bit Base model:moonshotai/kimi-k2-... Base model:quantized:moonshota... Conversational Custom code Kimi k2 Mlx Quantized Region:us Safetensors Sharded Tensorflow

Model Card on HF 🤗: https://huggingface.co/mlx-community/Kimi-K2-Thinking-3bit

Kimi K2 Thinking 3bit Benchmarks

LMSys ELO 1430 vs 1272 (so35) 12.4%

^nn.n% — How the model compares to the reference models: Anthropic Sonnet 3.5 ("so35"), GPT-4o ("gpt4o") or GPT-4 ("gpt4").

What is the LLM Explorer Rank (Score)

Kimi K2 Thinking 3bit Parameters and Internals

LLM Name	Kimi K2 Thinking 3bit
Repository 🤗	https://huggingface.co/mlx-community/Kimi-K2-Thinking-3bit
Base Model(s)	Kimi K2 Thinking moonshotai/Kimi-K2-Thinking
Model Size	1026.4b
Required VRAM	219.3 GB
Updated	2026-08-10
Maintainer	mlx-community
Model Type	kimi_k2
Model Files	5.4 GB: 1-of-91 5.3 GB: 2-of-91 5.0 GB: 3-of-91 4.9 GB: 4-of-91 5.0 GB: 5-of-91 5.0 GB: 6-of-91 4.9 GB: 7-of-91 5.0 GB: 8-of-91 5.0 GB: 9-of-91 4.9 GB: 10-of-91 5.0 GB: 11-of-91 5.0 GB: 12-of-91 4.9 GB: 13-of-91 5.0 GB: 14-of-91 5.0 GB: 15-of-91 4.9 GB: 16-of-91 5.0 GB: 17-of-91 5.0 GB: 18-of-91 4.9 GB: 19-of-91 5.0 GB: 20-of-91 5.0 GB: 21-of-91 4.9 GB: 22-of-91 5.0 GB: 23-of-91 5.0 GB: 24-of-91 4.9 GB: 25-of-91 5.0 GB: 26-of-91 5.0 GB: 27-of-91 4.9 GB: 28-of-91 5.0 GB: 29-of-91 5.0 GB: 30-of-91 4.9 GB: 31-of-91 5.0 GB: 32-of-91 5.0 GB: 33-of-91 4.9 GB: 34-of-91 5.0 GB: 35-of-91 5.0 GB: 36-of-91 4.9 GB: 37-of-91 5.0 GB: 38-of-91 5.0 GB: 39-of-91 4.9 GB: 40-of-91 5.0 GB: 41-of-91 5.0 GB: 42-of-91 4.9 GB: 43-of-91 5.0 GB: 44-of-91
Quantization Type	3bit
Model Architecture	DeepseekV3ForCausalLM
License	other
Context Length	262144
Model Max Length	262144
Transformers Version	4.51.3
Vocabulary Size	163840
Torch Data Type	bfloat16

Rank the Kimi K2 Thinking 3bit capabilities

Have you tried this model? Rate its performance — this feedback helps the ML community find the right model for their needs.

Instruction Following and Task Automation
Factuality and Completeness of Knowledge
Censorship and Alignment
Data Analysis and Insight Generation
Text Generation
Text Summarization and Feature Extraction
Code Generation
Multi-Language Support and Translation

Best Alternatives to Kimi K2 Thinking 3bit

Best Alternatives	Context / RAM	Downloads	Likes
Kimi K2 Thinking 4bit	256K / 151.2 GB	589	9
Kimi K2 Instruct 4bit	128K / 143.2 GB	5265	13
Kimi K2 Thinking	256K / 158.2 GB	5217	17
...i K2 Instruct 0905 Mlx DQ3 K M	256K / 204.6 GB	2234	12
Kimi K2 Instruct 0905 BF16	256K / 1399.1 GB	31	4
Kimi K2 Instruct BF16	128K / 1367.3 GB	20	4

Note: green Score (e.g. "73.2") means that the model is better than mlx-community/Kimi-K2-Thinking-3bit.