readygpu

Qwen 3.5 4B

4.7B parameters · 256k context · apache-2.0 · released 2026-02-27

chatvision

What to run it on in the US

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

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 3060 12GBQ8_0fits25 tok/s$36070.3
Intel Arc B580Q8_0fits26 tok/s$37069.7
Intel Arc A770 16GBQ8_0fits32 tok/s$40079.3
NVIDIA GeForce RTX 3080 10GBQ6_Kfits60 tok/s$469127.6
AMD Radeon RX 9060 XT 16GBQ8_0fits20 tok/s$54336.6
AMD Radeon RX 7800 XTQ8_0fits39 tok/s$59964.9
Intel Arc Pro B60Q8_0fits26 tok/s$65039.7
AMD Radeon RX 7600 XTQ8_0fits18 tok/s$69525.8
NVIDIA GeForce RTX 4070 SuperQ8_0fits35 tok/s$69950.6
NVIDIA GeForce RTX 4060 Ti 16GBQ8_0fits20 tok/s$77026.2
NVIDIA GeForce RTX 5060 Ti 16GBQ8_0fits32 tok/s$77240.8
NVIDIA GeForce RTX 5070Q8_0fits47 tok/s$77660.8
AMD Radeon RX 7900 XTQ8_0fits50 tok/s$80062.3
AMD Radeon RX 9070 XTQ8_0fits40 tok/s$81549.0
NVIDIA GeForce RTX 4070 Ti SuperQ8_0fits47 tok/s$1,01346.6
NVIDIA GeForce RTX 5070 TiQ8_0fits63 tok/s$1,14455.0
NVIDIA GeForce RTX 4080 SuperQ8_0fits52 tok/s$1,19943.1
AMD Radeon RX 7900 XTXQ8_0fits60 tok/s$1,53339.0
NVIDIA GeForce RTX 3090Q8_0fits66 tok/s$1,66039.6
AMD Radeon AI PRO R9700Q8_0fits40 tok/s$1,70023.5
NVIDIA GeForce RTX 5080Q8_0fits67 tok/s$1,73838.8
NVIDIA GeForce RTX 3090 TiQ8_0fits71 tok/s$2,06034.4
NVIDIA GeForce RTX 4090Q8_0fits71 tok/s$3,28021.6
NVIDIA GeForce RTX 5090Q8_0fits126 tok/s$7,20017.5
Mac Studio M4 Max 128GBQ8_0fits40 tok/s––
Mac mini M4 Pro 64GBQ8_0fits20 tok/s––
Framework Desktop Ryzen AI Max+ 395 128GBQ8_0fits16 tok/s––
Mac Studio M3 Ultra 512GBQ8_0fits60 tok/s––
NVIDIA DGX Spark 128GBQ8_0fits19 tok/s––
GMKtec EVO-X2 Ryzen AI Max+ 395 128GBQ8_0fits16 tok/s––
Mac Studio M3 Ultra 256GBQ8_0fits60 tok/s––
Mac mini M4 Pro 48GBQ8_0fits20 tok/s––
Mac Studio M4 Max 64GBQ8_0fits40 tok/s––
Mac Studio M3 Ultra 96GBQ8_0fits60 tok/s––
NVIDIA RTX PRO 6000 BlackwellQ8_0fits126 tok/s––
ASUS Ascent GX10 128GBQ8_0fits19 tok/s––
Framework Desktop Ryzen AI Max+ 395 64GBQ8_0fits16 tok/s––
Mac mini M4 32GBQ8_0fits9 tok/s––
Mac mini M4 16GBQ8_0fits9 tok/s––
NVIDIA Tesla P40Q8_0fits24 tok/s––

Memory needed

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

QuantFile@8k@32k@64k@128k
Q4_K_Mgguf2.6 GB3.6 GB4.5 GB5.7 GB8.0 GB
Q5_K_Mgguf2.9 GB4.0 GB4.9 GB6.0 GB8.4 GB
Q6_Kgguf3.3 GB4.3 GB5.2 GB6.4 GB8.8 GB
Q8_0gguf4.2 GB5.2 GB6.1 GB7.3 GB9.7 GB
mlx-8bitmlx4.8 GB5.8 GB6.7 GB7.9 GB10.3 GB

Run it

Copy a command for the runtime you use.

  • Q4_K_M
    • Ollamaollama run hf.co/unsloth/Qwen3.5-4B-GGUF:Q4_K_M
    • LM Studiolms get unsloth/Qwen3.5-4B-GGUF@q4_k_m
    • llama.cppllama-server -hf unsloth/Qwen3.5-4B-GGUF:Q4_K_M
  • Q5_K_M
    • Ollamaollama run hf.co/unsloth/Qwen3.5-4B-GGUF:Q5_K_M
    • LM Studiolms get unsloth/Qwen3.5-4B-GGUF@q5_k_m
    • llama.cppllama-server -hf unsloth/Qwen3.5-4B-GGUF:Q5_K_M
  • Q6_K
    • Ollamaollama run hf.co/unsloth/Qwen3.5-4B-GGUF:Q6_K
    • LM Studiolms get unsloth/Qwen3.5-4B-GGUF@q6_k
    • llama.cppllama-server -hf unsloth/Qwen3.5-4B-GGUF:Q6_K
  • Q8_0
    • Ollamaollama run hf.co/unsloth/Qwen3.5-4B-GGUF:Q8_0
    • LM Studiolms get unsloth/Qwen3.5-4B-GGUF@q8_0
    • llama.cppllama-server -hf unsloth/Qwen3.5-4B-GGUF:Q8_0
  • mlx-8bit
    • LM Studiolms get mlx-community/Qwen3.5-4B-8bit
    • MLXmlx_lm.server --model mlx-community/Qwen3.5-4B-8bit