readygpu

Gemma 4 E4B

8B parameters · 128k context · apache-2.0 · released 2026-03-02

chatvision

What to run it on in the US

Fully in memory at 32k 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 3090Q8_0fits63 tok/s––
NVIDIA GeForce RTX 5090Q8_0fits121 tok/s––
Mac Studio M4 Max 128GBQ8_0fits39 tok/s––
NVIDIA GeForce RTX 4090Q8_0fits68 tok/s––
NVIDIA GeForce RTX 5070 TiQ8_0fits61 tok/s––
Mac mini M4 Pro 64GBQ8_0fits19 tok/s––
NVIDIA GeForce RTX 5060 Ti 16GBQ8_0fits30 tok/s––
AMD Radeon RX 7900 XTXQ8_0fits58 tok/s––
Framework Desktop Ryzen AI Max+ 395 128GBQ8_0fits15 tok/s––
Mac Studio M3 Ultra 512GBQ8_0fits58 tok/s––
NVIDIA DGX Spark 128GBQ8_0fits19 tok/s––
NVIDIA GeForce RTX 3060 12GBQ8_0fits24 tok/s––
AMD Radeon RX 9070 XTQ8_0fits38 tok/s––
GMKtec EVO-X2 Ryzen AI Max+ 395 128GBQ8_0fits15 tok/s––
Mac Studio M3 Ultra 256GBQ8_0fits58 tok/s––
Mac mini M4 Pro 48GBQ8_0fits19 tok/s––
NVIDIA GeForce RTX 4060 Ti 16GBQ8_0fits20 tok/s––
NVIDIA GeForce RTX 5080Q8_0fits65 tok/s––
Mac Studio M4 Max 64GBQ8_0fits39 tok/s––
NVIDIA GeForce RTX 4070 Ti SuperQ8_0fits45 tok/s––
NVIDIA GeForce RTX 5070Q8_0fits45 tok/s––
AMD Radeon RX 9060 XT 16GBQ8_0fits19 tok/s––
Mac Studio M3 Ultra 96GBQ8_0fits58 tok/s––
NVIDIA GeForce RTX 4070 SuperQ8_0fits34 tok/s––
NVIDIA GeForce RTX 4080 SuperQ8_0fits50 tok/s––
NVIDIA RTX PRO 6000 BlackwellQ8_0fits121 tok/s––
AMD Radeon AI PRO R9700Q8_0fits38 tok/s––
AMD Radeon RX 7800 XTQ8_0fits37 tok/s––
AMD Radeon RX 7900 XTQ8_0fits48 tok/s––
ASUS Ascent GX10 128GBQ8_0fits19 tok/s––
Framework Desktop Ryzen AI Max+ 395 64GBQ8_0fits15 tok/s––
Intel Arc B580Q8_0fits25 tok/s––
Mac mini M4 32GBQ8_0fits9 tok/s––
NVIDIA GeForce RTX 3090 TiQ8_0fits68 tok/s––
AMD Radeon RX 7600 XTQ8_0fits17 tok/s––
Mac mini M4 16GBQ8_0fits9 tok/s––
NVIDIA GeForce RTX 3080 10GBQ6_Kfits59 tok/s––
Intel Arc Pro B60Q8_0fits25 tok/s––
NVIDIA Tesla P40Q8_0fits23 tok/s––
Intel Arc A770 16GBQ8_0fits31 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_Mgguf4.6 GB5.6 GB6.4 GB7.5 GB9.6 GB
mlx-4bitmlx4.8 GB5.8 GB6.6 GB7.7 GB9.8 GB
Q5_K_Mgguf5.1 GB6.1 GB6.9 GB8.0 GB10.1 GB
Q6_Kgguf6.6 GB7.6 GB8.4 GB9.4 GB11.6 GB
Q8_0gguf7.6 GB8.6 GB9.4 GB10.5 GB12.6 GB
mlx-8bitmlx8.3 GB9.3 GB10.1 GB11.1 GB13.3 GB

Run it

Copy a command for the runtime you use.

  • Q4_K_M
    • Ollamaollama run hf.co/unsloth/gemma-4-E4B-it-GGUF:Q4_K_M
    • LM Studiolms get unsloth/gemma-4-E4B-it-GGUF@q4_k_m
    • llama.cppllama-server -hf unsloth/gemma-4-E4B-it-GGUF:Q4_K_M
  • mlx-4bit
    • LM Studiolms get mlx-community/gemma-4-E4B-it-4bit
    • MLXmlx_lm.server --model mlx-community/gemma-4-E4B-it-4bit
  • Q5_K_M
    • Ollamaollama run hf.co/unsloth/gemma-4-E4B-it-GGUF:Q5_K_M
    • LM Studiolms get unsloth/gemma-4-E4B-it-GGUF@q5_k_m
    • llama.cppllama-server -hf unsloth/gemma-4-E4B-it-GGUF:Q5_K_M
  • Q6_K
    • Ollamaollama run hf.co/unsloth/gemma-4-E4B-it-GGUF:Q6_K
    • LM Studiolms get unsloth/gemma-4-E4B-it-GGUF@q6_k
    • llama.cppllama-server -hf unsloth/gemma-4-E4B-it-GGUF:Q6_K
  • Q8_0
    • Ollamaollama run hf.co/unsloth/gemma-4-E4B-it-GGUF:Q8_0
    • LM Studiolms get unsloth/gemma-4-E4B-it-GGUF@q8_0
    • llama.cppllama-server -hf unsloth/gemma-4-E4B-it-GGUF:Q8_0
  • mlx-8bit
    • LM Studiolms get mlx-community/gemma-4-E4B-it-8bit
    • MLXmlx_lm.server --model mlx-community/gemma-4-E4B-it-8bit