readygpu

Kimi K2 Thinking

1026.4B total, 32.9B active · 256k context · Modified MIT License · released 2025-11-04

chatcodingagenticreasoning

In short

Kimi K2 Thinking is an open model from Moonshot AI with 1026.4B parameters (32.9B active per token). At mlx-3bit, Kimi K2 Thinking needs about 420 GB of memory at 8k context and 421 GB at 32k, counting the weights, the KV cache and runtime buffers. Nothing we price in Canada runs Kimi K2 Thinking fully in memory at 32k context at a usable speed; see the table for what comes closest.

Updated . Speeds are estimates; how they are worked out.

What to run it on in Canada

Fully in memory at 32k context, priced at retailers today.

Nothing we price holds this model fully in memory.

All hardware that runs it

Everything we price, best value first. Two cards split the model by layers.

Nothing we price can run it at this context.

Memory needed

Weights, KV cache and runtime buffers at each context length (KV cache at fp16).

QuantFile@8k@32k@64k@128k
mlx-3bitmlx418.5 GB419.7 GB421.5 GB423.8 GB428.5 GB
Q4_K_Mgguf578.6 GB579.9 GB581.6 GB583.9 GB588.6 GB
Q5_K_Mgguf678.7 GB680.0 GB681.7 GB684.1 GB688.7 GB
Q6_Kgguf785.0 GB786.3 GB788.1 GB790.4 GB795.0 GB
Q8_0gguf1016.1 GB1017.4 GB1019.2 GB1021.5 GB1026.2 GB

Run it

Copy a command for the runtime you use.

  • mlx-3bit
    • LM Studiolms get mlx-community/Kimi-K2-Thinking-3bit
    • MLXmlx_lm.server --model mlx-community/Kimi-K2-Thinking-3bit
  • Q4_K_M
    • Ollamaollama run hf.co/unsloth/Kimi-K2-Thinking-GGUF:Q4_K_M
    • LM Studiolms get unsloth/Kimi-K2-Thinking-GGUF@q4_k_m
    • llama.cppllama-server -hf unsloth/Kimi-K2-Thinking-GGUF:Q4_K_M
  • Q5_K_M
    • Ollamaollama run hf.co/unsloth/Kimi-K2-Thinking-GGUF:Q5_K_M
    • LM Studiolms get unsloth/Kimi-K2-Thinking-GGUF@q5_k_m
    • llama.cppllama-server -hf unsloth/Kimi-K2-Thinking-GGUF:Q5_K_M
  • Q6_K
    • Ollamaollama run hf.co/unsloth/Kimi-K2-Thinking-GGUF:Q6_K
    • LM Studiolms get unsloth/Kimi-K2-Thinking-GGUF@q6_k
    • llama.cppllama-server -hf unsloth/Kimi-K2-Thinking-GGUF:Q6_K
  • Q8_0
    • Ollamaollama run hf.co/unsloth/Kimi-K2-Thinking-GGUF:Q8_0
    • LM Studiolms get unsloth/Kimi-K2-Thinking-GGUF@q8_0
    • llama.cppllama-server -hf unsloth/Kimi-K2-Thinking-GGUF:Q8_0

Notes

1T-parameter MoE (32B active) shipped as native INT4 (QAT) weights in compressed-tensors format; recommended engines are vLLM, SGLang and KTransformers. Licence is a Modified MIT whose terms are not reproduced in the card, so check the LICENSE file before commercial use.

Questions

How much VRAM does Kimi K2 Thinking need?
At mlx-3bit, Kimi K2 Thinking needs about 420 GB of memory at 8k context and 421 GB at 32k, counting the weights, the KV cache and runtime buffers.
Can I run Kimi K2 Thinking on a 16 GB GPU?
No: it does not fit in 16 GB at 32k context, even with system RAM.
Can I run Kimi K2 Thinking on a 24 GB GPU?
No: it does not fit in 24 GB at 32k context, even with system RAM.
How do I run Kimi K2 Thinking locally?
With Ollama: ollama run hf.co/unsloth/Kimi-K2-Thinking-GGUF:Q4_K_M (the Q4_K_M build). LM Studio and llama.cpp commands are on this page.
What licence is Kimi K2 Thinking released under?
Modified MIT License.