readygpu

DeepSeek R1 Distill Qwen 32B

32.8B parameters · 128k context · mit · released 2025-01-20

chatreasoning

What to run it on in India

Fully in memory at 64k 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
Mac Studio M4 Max 128GBQ8_0fits7 tok/s––
Mac mini M4 Pro 64GBQ6_Kfits4 tok/s––
Framework Desktop Ryzen AI Max+ 395 128GBQ8_0fits3 tok/s––
Mac Studio M3 Ultra 512GBQ8_0fits10 tok/s––
NVIDIA DGX Spark 128GBQ8_0fits3 tok/s––
GMKtec EVO-X2 Ryzen AI Max+ 395 128GBQ8_0fits3 tok/s––
Mac Studio M3 Ultra 256GBQ8_0fits10 tok/s––
Mac mini M4 Pro 48GBQ4_K_Mfits5 tok/s––
Mac Studio M4 Max 64GBQ6_Kfits8 tok/s––
Mac Studio M3 Ultra 96GBQ8_0fits10 tok/s––
NVIDIA RTX PRO 6000 BlackwellQ8_0fits21 tok/s––
ASUS Ascent GX10 128GBQ8_0fits3 tok/s––
Framework Desktop Ryzen AI Max+ 395 64GBQ6_Kfits3 tok/s––
NVIDIA GeForce RTX 3090Q4_K_Moffload4 tok/s––
NVIDIA GeForce RTX 5090Q4_K_Moffload9 tok/s––
NVIDIA GeForce RTX 4090Q4_K_Moffload4 tok/s––
NVIDIA GeForce RTX 5070 TiQ4_K_Moffload2 tok/s––
NVIDIA GeForce RTX 5060 Ti 16GBQ4_K_Moffload2 tok/s––
AMD Radeon RX 7900 XTXQ4_K_Moffload3 tok/s––
NVIDIA GeForce RTX 3060 12GBQ4_K_Moffload2 tok/s––
AMD Radeon RX 9070 XTQ4_K_Moffload2 tok/s––
NVIDIA GeForce RTX 4060 Ti 16GBQ4_K_Moffload2 tok/s––
NVIDIA GeForce RTX 5080Q4_K_Moffload2 tok/s––
NVIDIA GeForce RTX 4070 Ti SuperQ4_K_Moffload2 tok/s––
NVIDIA GeForce RTX 5070Q4_K_Moffload2 tok/s––
AMD Radeon RX 9060 XT 16GBQ4_K_Moffload2 tok/s––
NVIDIA GeForce RTX 4070 SuperQ4_K_Moffload2 tok/s––
NVIDIA GeForce RTX 4080 SuperQ4_K_Moffload2 tok/s––
AMD Radeon AI PRO R9700Q4_K_Moffload6 tok/s––
AMD Radeon RX 7800 XTQ4_K_Moffload2 tok/s––
AMD Radeon RX 7900 XTQ4_K_Moffload3 tok/s––
Intel Arc B580Q4_K_Moffload2 tok/s––
NVIDIA GeForce RTX 3090 TiQ4_K_Moffload4 tok/s––
AMD Radeon RX 7600 XTQ4_K_Moffload2 tok/s––
Intel Arc Pro B60Q4_K_Moffload3 tok/s––
NVIDIA Tesla P40Q4_K_Moffload3 tok/s––
Intel Arc A770 16GBQ4_K_Moffload2 tok/s––

3 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_Mgguf18.5 GB21.2 GB27.4 GB35.6 GB51.9 GB
mlx-4bitmlx17.2 GB19.9 GB26.1 GB34.2 GB50.6 GB
Q5_K_Mgguf21.7 GB24.4 GB30.6 GB38.7 GB55.1 GB
Q6_Kgguf25.0 GB27.8 GB33.9 GB42.1 GB58.5 GB
Q8_0gguf32.4 GB35.2 GB41.3 GB49.5 GB65.9 GB

Run it

Copy a command for the runtime you use.

  • Q4_K_M
    • Ollamaollama run hf.co/unsloth/DeepSeek-R1-Distill-Qwen-32B-GGUF:Q4_K_M
    • LM Studiolms get unsloth/DeepSeek-R1-Distill-Qwen-32B-GGUF@q4_k_m
    • llama.cppllama-server -hf unsloth/DeepSeek-R1-Distill-Qwen-32B-GGUF:Q4_K_M
  • mlx-4bit
    • LM Studiolms get mlx-community/DeepSeek-R1-Distill-Qwen-32B-4bit
    • MLXmlx_lm.server --model mlx-community/DeepSeek-R1-Distill-Qwen-32B-4bit
  • Q5_K_M
    • Ollamaollama run hf.co/unsloth/DeepSeek-R1-Distill-Qwen-32B-GGUF:Q5_K_M
    • LM Studiolms get unsloth/DeepSeek-R1-Distill-Qwen-32B-GGUF@q5_k_m
    • llama.cppllama-server -hf unsloth/DeepSeek-R1-Distill-Qwen-32B-GGUF:Q5_K_M
  • Q6_K
    • Ollamaollama run hf.co/unsloth/DeepSeek-R1-Distill-Qwen-32B-GGUF:Q6_K
    • LM Studiolms get unsloth/DeepSeek-R1-Distill-Qwen-32B-GGUF@q6_k
    • llama.cppllama-server -hf unsloth/DeepSeek-R1-Distill-Qwen-32B-GGUF:Q6_K
  • Q8_0
    • Ollamaollama run hf.co/unsloth/DeepSeek-R1-Distill-Qwen-32B-GGUF:Q8_0
    • LM Studiolms get unsloth/DeepSeek-R1-Distill-Qwen-32B-GGUF@q8_0
    • llama.cppllama-server -hf unsloth/DeepSeek-R1-Distill-Qwen-32B-GGUF:Q8_0