RUN THIS LLM
Search local LLM hardware requirements
Gemma 4 26B MoE
Google · 26B MoE · Vision
Gemma 4 sparse MoE with 26B total params and only 4B active per token. Flagship-level multimodal quality at efficient inference cost.
VRAM Requirements
| Quantization | VRAM |
|---|---|
| Q4_K_M (smallest) | 15.6 GB |
| Q8_0 (balanced) | 28.6 GB |
| FP16 (full quality) | 52 GB |
Specifications
- Parameters: 26B MoE (4B active per token)
- Category: Vision
- Max context: 128K tokens
- System RAM: 24 GB minimum
- HuggingFace: google/gemma-4-26B-A4B-it
Benchmarks
- MMLU: 83 — General knowledge & reasoning (5-shot)
- HumanEval: 77 — Code generation (pass@1)
- MATH: 88 — Competition-level math reasoning
Loading interactive analysis...