readygpu

Qwen3 32B

32.8B parameters · 40k context · apache-2.0 · released 2025-04-27

chatcodingreasoning

What to run it on in Germany

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
Mac Studio M4 Max 128GBQ8_0fits5 tok/s––
Framework Desktop Ryzen AI Max+ 395 128GBQ8_0fits2 tok/s––
Mac Studio M3 Ultra 512GBQ8_0fits8 tok/s––
NVIDIA DGX Spark 128GBQ8_0fits2 tok/s––
GMKtec EVO-X2 Ryzen AI Max+ 395 128GBQ8_0fits2 tok/s––
Mac Studio M3 Ultra 256GBQ8_0fits8 tok/s––
Mac Studio M3 Ultra 96GBQ8_0fits8 tok/s––
NVIDIA RTX PRO 6000 BlackwellQ8_0fits16 tok/s––
ASUS Ascent GX10 128GBQ8_0fits2 tok/s––
NVIDIA GeForce RTX 5090Q4_K_Moffload2 tok/s––
AMD Radeon AI PRO R9700Q4_K_Moffload2 tok/s––

29 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
Q4_K_Mgguf18.4 GB21.2 GB27.3 GB
mlx-4bitmlx17.2 GB19.9 GB26.1 GB
Q5_K_Mgguf21.6 GB24.4 GB30.5 GB
Q6_Kgguf25.0 GB27.8 GB33.9 GB
mlx-8bitmlx32.4 GB35.2 GB41.3 GB
Q8_0gguf32.4 GB35.2 GB41.3 GB

Run it

Copy a command for the runtime you use.

  • Q4_K_M
    • Ollamaollama run hf.co/unsloth/Qwen3-32B-GGUF:Q4_K_M
    • LM Studiolms get unsloth/Qwen3-32B-GGUF@q4_k_m
    • llama.cppllama-server -hf unsloth/Qwen3-32B-GGUF:Q4_K_M
  • mlx-4bit
    • LM Studiolms get mlx-community/Qwen3-32B-4bit
    • MLXmlx_lm.server --model mlx-community/Qwen3-32B-4bit
  • Q5_K_M
    • Ollamaollama run hf.co/unsloth/Qwen3-32B-GGUF:Q5_K_M
    • LM Studiolms get unsloth/Qwen3-32B-GGUF@q5_k_m
    • llama.cppllama-server -hf unsloth/Qwen3-32B-GGUF:Q5_K_M
  • Q6_K
    • Ollamaollama run hf.co/unsloth/Qwen3-32B-GGUF:Q6_K
    • LM Studiolms get unsloth/Qwen3-32B-GGUF@q6_k
    • llama.cppllama-server -hf unsloth/Qwen3-32B-GGUF:Q6_K
  • mlx-8bit
    • LM Studiolms get mlx-community/Qwen3-32B-8bit
    • MLXmlx_lm.server --model mlx-community/Qwen3-32B-8bit
  • Q8_0
    • Ollamaollama run hf.co/unsloth/Qwen3-32B-GGUF:Q8_0
    • LM Studiolms get unsloth/Qwen3-32B-GGUF@q8_0
    • llama.cppllama-server -hf unsloth/Qwen3-32B-GGUF:Q8_0