← CanIRunLocal · All hardware · All models

Can I run Gemma 4 26B-A4B (MoE) locally?

Google · 25.8B parameters (3.8B active) · chat, reasoning, coding, vision.

Ollama weight files: Q4_K_M 18 GB (gemma4:26b-a4b-it-q4_K_M); Q8_0 28 GB (gemma4:26b-a4b-it-q8_0).

Yes on 48 machines in this list. The smallest is RTX 3080 10GB: Q4_K_M, about 18.5 GB, roughly 15 tokens per second, via LM Studio. lms get unsloth/gemma-4-26B-A4B-it-GGUF@Q4_K_M

• DGX Spark 128GB: Q8_0, ~29.6 GB, fits in fast memory, ~38 tok/s — ollama run gemma4:26b-a4b-it-q8_0
• Mac Studio M5 Max 128GB: Q8_0, ~29.6 GB, fits in fast memory, ~60 tok/s — ollama run gemma4:26b-a4b-it-q8_0
• Mac Studio M5 Ultra 256GB: Q8_0, ~29.6 GB, fits in fast memory, ~116 tok/s — ollama run gemma4:26b-a4b-it-q8_0

Setup on DGX Spark 128GB (NVIDIA, ASUS, Dell, Lenovo, HP, MSI, Acer, Gigabyte)

• Ollama: ollama run gemma4:26b-a4b-it-q8_0
• LM Studio: lms get unsloth/gemma-4-26B-A4B-it-GGUF@Q8_0
• llama.cpp: llama-server -hf unsloth/gemma-4-26B-A4B-it-GGUF:Q8_0 -c 8192
• vLLM: vllm serve google/gemma-4-26b-a4b-it --max-model-len 8192 --gpu-memory-utilization 0.9

Hardware that fits

Estimates, not benchmarks. Model sizes are ollama.com download sizes; memory needed adds headroom for ~8k tokens of context; speeds are rough figures from memory bandwidth. Data updated 2026-10-18.

A project by OnDevice Labs · llms.txt