Model · Google

Gemma 4 26B-A4B

25.8B MoE · Q4_K_M · 15.8 GB · open weights · curated for general, research · repo created 2026-03-11

Runs on a 24 GB card at 8K context: 17.5 GB. At 32K it needs 18.7 GB there and still runs.

What the model needs: weights, context cache and runtime overhead at four reference contexts. The total is the model's own; what differs per machine is the usable memory it has to fit into — VRAM on a graphics card, unified memory on a Mac.

Cache shape from the model's config: 5 of 30 layers cache the full context with their own global heads, K and V for 2 KV heads × 512 dimensions per token at f16; 25 layers see a 1,024-token sliding window and cache only that, those with 8 KV heads × 256 dimensions. Multiply heads by dimensions by two bytes by tokens per layer kind and you have the cache column.

All memory figures are estimates: measured quantized file size + computed context memory + runtime overhead, with a 12% safety margin on your hardware. Real usage varies with runtime version and settings.

Every tracked GPU and Mac at 8K and 32K context with a full-precision (f16) cache — the same engine as the finder and the GPU pages.

Runs on 11 of 32 tracked devices at 8K context, easily or with ordinary headroom.

GGUF is the quantized single-file format local runtimes load (Ollama, LM Studio, llama.cpp); our memory figures are measured from the linked GGUF file. The original repo holds full-precision safetensors — a much larger download meant for GPUs with far more memory.

More from Google

• Gemma 4 E4Bmore popular ▲
• Gemma 4 31Bmore popular ▲
• Gemma 4 12Bmore popular ▲
• Gemma 3 27Bmore popular ▲

Every tracked model →