readygpu

Kimi Linear 48B-A3B

49.1B total, 3B active · 1024k context · mit · released 2025-10-30

chatcoding

What to run it on in the UK

Fully in memory at 128k context, priced at retailers today.

Something we price holds it fully in memory, but none reaches 15 tok/s, our bar for recommending it. The table below has the speeds.

All hardware that runs it

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

HardwareQuantFitsDecodePricetok/s per 1,000
NVIDIA GeForce RTX 5090Q4_K_Mfits227 tok/s––
Mac Studio M4 Max 128GBQ8_0fits50 tok/s––
Mac mini M4 Pro 64GBQ6_Kfits30 tok/s––
Framework Desktop Ryzen AI Max+ 395 128GBQ8_0fits20 tok/s––
Mac Studio M3 Ultra 512GBQ8_0fits75 tok/s––
NVIDIA DGX Spark 128GBQ8_0fits24 tok/s––
GMKtec EVO-X2 Ryzen AI Max+ 395 128GBQ8_0fits20 tok/s––
Mac Studio M3 Ultra 256GBQ8_0fits75 tok/s––
Mac mini M4 Pro 48GBQ5_K_Mfits33 tok/s––
Mac Studio M4 Max 64GBQ6_Kfits60 tok/s––
Mac Studio M3 Ultra 96GBQ8_0fits75 tok/s––
NVIDIA RTX PRO 6000 BlackwellQ8_0fits155 tok/s––
AMD Radeon AI PRO R9700Q4_K_Mfits72 tok/s––
ASUS Ascent GX10 128GBQ8_0fits24 tok/s––
Framework Desktop Ryzen AI Max+ 395 64GBQ6_Kfits24 tok/s––
NVIDIA GeForce RTX 3090Q4_K_Moffload35 tok/s––
NVIDIA GeForce RTX 4090Q4_K_Moffload36 tok/s––
NVIDIA GeForce RTX 5070 TiQ4_K_Moffload19 tok/s––
NVIDIA GeForce RTX 5060 Ti 16GBQ4_K_Moffload18 tok/s––
AMD Radeon RX 7900 XTXQ4_K_Moffload33 tok/s––
NVIDIA GeForce RTX 3060 12GBQ4_K_Moffload15 tok/s––
AMD Radeon RX 9070 XTQ4_K_Moffload18 tok/s––
NVIDIA GeForce RTX 4060 Ti 16GBQ4_K_Moffload17 tok/s––
NVIDIA GeForce RTX 5080Q4_K_Moffload20 tok/s––
NVIDIA GeForce RTX 4070 Ti SuperQ4_K_Moffload19 tok/s––
NVIDIA GeForce RTX 5070Q4_K_Moffload16 tok/s––
AMD Radeon RX 9060 XT 16GBQ4_K_Moffload16 tok/s––
NVIDIA GeForce RTX 4070 SuperQ4_K_Moffload15 tok/s––
NVIDIA GeForce RTX 4080 SuperQ4_K_Moffload19 tok/s––
AMD Radeon RX 7800 XTQ4_K_Moffload18 tok/s––
AMD Radeon RX 7900 XTQ4_K_Moffload24 tok/s––
Intel Arc B580Q4_K_Moffload14 tok/s––
NVIDIA GeForce RTX 3090 TiQ4_K_Moffload36 tok/s––
AMD Radeon RX 7600 XTQ4_K_Moffload16 tok/s––
NVIDIA GeForce RTX 3080 10GBQ4_K_Moffload15 tok/s––
Intel Arc Pro B60Q4_K_Moffload23 tok/s––
NVIDIA Tesla P40Q4_K_Moffload25 tok/s––
Intel Arc A770 16GBQ4_K_Moffload16 tok/s––

2 other cards and machines we price cannot hold 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
Q4_K_Mgguf28.0 GB28.8 GB29.2 GB29.6 GB30.5 GB
mlx-4bitmlx25.7 GB26.6 GB26.9 GB27.3 GB28.2 GB
Q5_K_Mgguf32.7 GB33.5 GB33.9 GB34.3 GB35.2 GB
Q6_Kgguf37.7 GB38.5 GB38.9 GB39.3 GB40.2 GB
mlx-8bitmlx48.6 GB49.5 GB49.8 GB50.2 GB51.1 GB
Q8_0gguf48.7 GB49.5 GB49.8 GB50.3 GB51.1 GB

Run it

Copy a command for the runtime you use.

  • Q4_K_M
    • Ollamaollama run hf.co/bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF:Q4_K_M
    • LM Studiolms get bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF@q4_k_m
    • llama.cppllama-server -hf bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF:Q4_K_M
  • mlx-4bit
    • LM Studiolms get mlx-community/Kimi-Linear-48B-A3B-Instruct-4bit
    • MLXmlx_lm.server --model mlx-community/Kimi-Linear-48B-A3B-Instruct-4bit
  • Q5_K_M
    • Ollamaollama run hf.co/bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF:Q5_K_M
    • LM Studiolms get bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF@q5_k_m
    • llama.cppllama-server -hf bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF:Q5_K_M
  • Q6_K
    • Ollamaollama run hf.co/bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF:Q6_K
    • LM Studiolms get bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF@q6_k
    • llama.cppllama-server -hf bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF:Q6_K
  • mlx-8bit
    • LM Studiolms get mlx-community/Kimi-Linear-48B-A3B-Instruct-8bit
    • MLXmlx_lm.server --model mlx-community/Kimi-Linear-48B-A3B-Instruct-8bit
  • Q8_0
    • Ollamaollama run hf.co/bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF:Q8_0
    • LM Studiolms get bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF@q8_0
    • llama.cppllama-server -hf bartowski/moonshotai_Kimi-Linear-48B-A3B-Instruct-GGUF:Q8_0