readygpu

Gemma 4 26B-A4B

25.8B total, 4B active · 256k context · apache-2.0 · released 2026-03-11

chatcodingagenticreasoningvision

What to run it on in the US

Fully in memory at 8k 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 3090UD-Q6_Kfits88 tok/s––
NVIDIA GeForce RTX 5090Q8_0fits147 tok/s––
Mac Studio M4 Max 128GBQ8_0fits47 tok/s––
NVIDIA GeForce RTX 4090UD-Q6_Kfits95 tok/s––
Mac mini M4 Pro 64GBQ8_0fits23 tok/s––
AMD Radeon RX 7900 XTXUD-Q6_Kfits80 tok/s––
Framework Desktop Ryzen AI Max+ 395 128GBQ8_0fits19 tok/s––
Mac Studio M3 Ultra 512GBQ8_0fits70 tok/s––
NVIDIA DGX Spark 128GBQ8_0fits22 tok/s––
GMKtec EVO-X2 Ryzen AI Max+ 395 128GBQ8_0fits19 tok/s––
Mac Studio M3 Ultra 256GBQ8_0fits70 tok/s––
Mac mini M4 Pro 48GBQ8_0fits23 tok/s––
Mac Studio M4 Max 64GBQ8_0fits47 tok/s––
Mac Studio M3 Ultra 96GBQ8_0fits70 tok/s––
NVIDIA RTX PRO 6000 BlackwellQ8_0fits147 tok/s––
AMD Radeon AI PRO R9700Q8_0fits47 tok/s––
AMD Radeon RX 7900 XTUD-Q4_K_Mfits88 tok/s––
ASUS Ascent GX10 128GBQ8_0fits22 tok/s––
Framework Desktop Ryzen AI Max+ 395 64GBQ8_0fits19 tok/s––
Mac mini M4 32GBUD-Q5_K_Mfits13 tok/s––
NVIDIA GeForce RTX 3090 TiUD-Q6_Kfits95 tok/s––
Intel Arc Pro B60UD-Q6_Kfits32 tok/s––
NVIDIA Tesla P40UD-Q6_Kfits32 tok/s––
NVIDIA GeForce RTX 5070 TiUD-Q4_K_Moffload60 tok/s––
NVIDIA GeForce RTX 5060 Ti 16GBUD-Q4_K_Moffload40 tok/s––
NVIDIA GeForce RTX 3060 12GBUD-Q4_K_Moffload22 tok/s––
AMD Radeon RX 9070 XTUD-Q4_K_Moffload44 tok/s––
NVIDIA GeForce RTX 4060 Ti 16GBUD-Q4_K_Moffload29 tok/s––
NVIDIA GeForce RTX 5080UD-Q4_K_Moffload62 tok/s––
NVIDIA GeForce RTX 4070 Ti SuperUD-Q4_K_Moffload51 tok/s––
NVIDIA GeForce RTX 5070UD-Q4_K_Moffload25 tok/s––
AMD Radeon RX 9060 XT 16GBUD-Q4_K_Moffload28 tok/s––
NVIDIA GeForce RTX 4070 SuperUD-Q4_K_Moffload24 tok/s––
NVIDIA GeForce RTX 4080 SuperUD-Q4_K_Moffload54 tok/s––
AMD Radeon RX 7800 XTUD-Q4_K_Moffload43 tok/s––
Intel Arc B580UD-Q4_K_Moffload19 tok/s––
AMD Radeon RX 7600 XTUD-Q4_K_Moffload26 tok/s––
NVIDIA GeForce RTX 3080 10GBUD-Q4_K_Moffload21 tok/s––
Intel Arc A770 16GBUD-Q4_K_Moffload35 tok/s––

1 other card or machine 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
UD-Q4_K_Mgguf15.8 GB16.9 GB17.5 GB18.3 GB19.9 GB
mlx-4bitmlx14.3 GB15.4 GB16.0 GB16.8 GB18.4 GB
UD-Q5_K_Mgguf19.7 GB20.8 GB21.4 GB22.2 GB23.8 GB
UD-Q6_Kgguf21.6 GB22.7 GB23.3 GB24.1 GB25.7 GB
Q8_0gguf25.0 GB26.1 GB26.7 GB27.5 GB29.2 GB
mlx-8bitmlx26.0 GB27.1 GB27.7 GB28.6 GB30.2 GB

Run it

Copy a command for the runtime you use.

  • UD-Q4_K_M
    • Ollamaollama run hf.co/unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q4_K_M
    • LM Studiolms get unsloth/gemma-4-26B-A4B-it-GGUF@ud-q4_k_m
    • llama.cppllama-server -hf unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q4_K_M
  • mlx-4bit
    • LM Studiolms get mlx-community/gemma-4-26B-A4B-it-4bit
    • MLXmlx_lm.server --model mlx-community/gemma-4-26B-A4B-it-4bit
  • UD-Q5_K_M
    • Ollamaollama run hf.co/unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q5_K_M
    • LM Studiolms get unsloth/gemma-4-26B-A4B-it-GGUF@ud-q5_k_m
    • llama.cppllama-server -hf unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q5_K_M
  • UD-Q6_K
    • Ollamaollama run hf.co/unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q6_K
    • LM Studiolms get unsloth/gemma-4-26B-A4B-it-GGUF@ud-q6_k
    • llama.cppllama-server -hf unsloth/gemma-4-26B-A4B-it-GGUF:UD-Q6_K
  • Q8_0
    • Ollamaollama run hf.co/unsloth/gemma-4-26B-A4B-it-GGUF:Q8_0
    • LM Studiolms get unsloth/gemma-4-26B-A4B-it-GGUF@q8_0
    • llama.cppllama-server -hf unsloth/gemma-4-26B-A4B-it-GGUF:Q8_0
  • mlx-8bit
    • LM Studiolms get mlx-community/gemma-4-26B-A4B-it-8bit
    • MLXmlx_lm.server --model mlx-community/gemma-4-26B-A4B-it-8bit