readygpu

Gemma 4 12B

12B parameters · 256k context · apache-2.0 · released 2026-05-23

chatcodingvision

What to run it on in the US

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 3090Q8_0fits38 tok/s––
NVIDIA GeForce RTX 5090Q8_0fits73 tok/s––
Mac Studio M4 Max 128GBQ8_0fits23 tok/s––
NVIDIA GeForce RTX 4090Q8_0fits41 tok/s––
NVIDIA GeForce RTX 5070 TiQ6_Kfits45 tok/s––
Mac mini M4 Pro 64GBQ8_0fits12 tok/s––
NVIDIA GeForce RTX 5060 Ti 16GBQ6_Kfits23 tok/s––
AMD Radeon RX 7900 XTXQ8_0fits35 tok/s––
Framework Desktop Ryzen AI Max+ 395 128GBQ8_0fits9 tok/s––
Mac Studio M3 Ultra 512GBQ8_0fits35 tok/s––
NVIDIA DGX Spark 128GBQ8_0fits11 tok/s––
NVIDIA GeForce RTX 3060 12GBQ4_K_Mfits23 tok/s––
AMD Radeon RX 9070 XTQ6_Kfits29 tok/s––
GMKtec EVO-X2 Ryzen AI Max+ 395 128GBQ8_0fits9 tok/s––
Mac Studio M3 Ultra 256GBQ8_0fits35 tok/s––
Mac mini M4 Pro 48GBQ8_0fits12 tok/s––
NVIDIA GeForce RTX 4060 Ti 16GBQ6_Kfits15 tok/s––
NVIDIA GeForce RTX 5080Q6_Kfits49 tok/s––
Mac Studio M4 Max 64GBQ8_0fits23 tok/s––
NVIDIA GeForce RTX 4070 Ti SuperQ6_Kfits34 tok/s––
NVIDIA GeForce RTX 5070Q4_K_Mfits43 tok/s––
AMD Radeon RX 9060 XT 16GBQ6_Kfits14 tok/s––
Mac Studio M3 Ultra 96GBQ8_0fits35 tok/s––
NVIDIA GeForce RTX 4070 SuperQ4_K_Mfits33 tok/s––
NVIDIA GeForce RTX 4080 SuperQ6_Kfits37 tok/s––
NVIDIA RTX PRO 6000 BlackwellQ8_0fits73 tok/s––
AMD Radeon AI PRO R9700Q8_0fits23 tok/s––
AMD Radeon RX 7800 XTQ6_Kfits28 tok/s––
AMD Radeon RX 7900 XTQ8_0fits29 tok/s––
ASUS Ascent GX10 128GBQ8_0fits11 tok/s––
Framework Desktop Ryzen AI Max+ 395 64GBQ8_0fits9 tok/s––
Intel Arc B580Q4_K_Mfits24 tok/s––
Mac mini M4 32GBQ8_0fits5 tok/s––
NVIDIA GeForce RTX 3090 TiQ8_0fits41 tok/s––
AMD Radeon RX 7600 XTQ6_Kfits13 tok/s––
Mac mini M4 16GBQ4_K_Mfits8 tok/s––
Intel Arc Pro B60Q8_0fits15 tok/s––
NVIDIA Tesla P40Q8_0fits14 tok/s––
Intel Arc A770 16GBQ6_Kfits23 tok/s––
NVIDIA GeForce RTX 3080 10GBQ4_K_Moffload26 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_Mgguf6.6 GB7.8 GB8.3 GB9.0 GB10.4 GB
mlx-4bitmlx6.3 GB7.5 GB8.0 GB8.7 GB10.0 GB
Q5_K_Mgguf7.8 GB9.0 GB9.5 GB10.2 GB11.6 GB
Q6_Kgguf9.1 GB10.3 GB10.8 GB11.5 GB12.9 GB
Q8_0gguf11.8 GB13.0 GB13.5 GB14.2 GB15.6 GB
mlx-8bitmlx11.8 GB13.0 GB13.5 GB14.2 GB15.6 GB

Run it

Copy a command for the runtime you use.

  • Q4_K_M
    • Ollamaollama run hf.co/unsloth/gemma-4-12b-it-GGUF:Q4_K_M
    • LM Studiolms get unsloth/gemma-4-12b-it-GGUF@q4_k_m
    • llama.cppllama-server -hf unsloth/gemma-4-12b-it-GGUF:Q4_K_M
  • mlx-4bit
    • LM Studiolms get mlx-community/gemma-4-12B-it-4bit
    • MLXmlx_lm.server --model mlx-community/gemma-4-12B-it-4bit
  • Q5_K_M
    • Ollamaollama run hf.co/unsloth/gemma-4-12b-it-GGUF:Q5_K_M
    • LM Studiolms get unsloth/gemma-4-12b-it-GGUF@q5_k_m
    • llama.cppllama-server -hf unsloth/gemma-4-12b-it-GGUF:Q5_K_M
  • Q6_K
    • Ollamaollama run hf.co/unsloth/gemma-4-12b-it-GGUF:Q6_K
    • LM Studiolms get unsloth/gemma-4-12b-it-GGUF@q6_k
    • llama.cppllama-server -hf unsloth/gemma-4-12b-it-GGUF:Q6_K
  • Q8_0
    • Ollamaollama run hf.co/unsloth/gemma-4-12b-it-GGUF:Q8_0
    • LM Studiolms get unsloth/gemma-4-12b-it-GGUF@q8_0
    • llama.cppllama-server -hf unsloth/gemma-4-12b-it-GGUF:Q8_0
  • mlx-8bit
    • LM Studiolms get mlx-community/gemma-4-12B-it-8bit
    • MLXmlx_lm.server --model mlx-community/gemma-4-12B-it-8bit