Which Gemma 4 Model Should You Run?
Tell us what you mainly want to do and what hardware you have — we'll filter to what actually fits, then rank the rest using our published benchmark and comparison data.
Gemma 4 12B
✅ Comfortable · Q4_0 (4-bit)Weights: ~6.7 GB
No official Arena Elo is published yet, but its reasoning scores land as a capable daily driver — GPQA Diamond 78.8%, τ2-bench 69.0% — clearly ahead of E4B and within a few points of 26B A4B on several tasks.
Grab the official QAT build (~7.2 GB) for better quality at essentially the same size — see the QAT guide.
Gemma 4 E4B
✅ Comfortable · SFP8 (8-bit)Weights: ~7.5 GB
The everyday laptop default — "best balance of quality and speed on 8–16 GB hardware" — though its benchmark scores (MMLU Pro 69.4%) trail the three bigger dense/MoE models by a wide margin.
How this works
This tool runs two passes. First, it filters out any Gemma 4 size that doesn't fit your hardware at all, using the same judgement engine as our hardware compatibility checker. Second, it ranks whatever's left using benchmark scores and use-case notes from our benchmarks breakdown, features & use cases guide, and 12B vs 26B A4B comparison.
Where an article doesn't publish a number for a given model and use case — E2B has no coding or vision benchmark, for instance — the recommendation says so directly instead of inventing a score.
FAQ
Which Gemma 4 model is best overall?
Gemma 4 31B posts the best benchmark scores across reasoning, coding, and vision, and it has the highest real-world preference in blind testing (Arena AI Elo 1452). But 'best' only matters if your hardware runs it comfortably — 26B A4B gets close on quality at a fraction of the compute, and 12B or E4B are better fits for smaller machines.
Gemma 4 12B or 26B A4B — which one should I run?
26B A4B wins nearly every benchmark, sometimes by double digits, but needs roughly 2.3x the memory and has no native audio support. 12B is the model built for a 16 GB machine, has a mature QAT path, and is the only option if you need audio input. If you have 24 GB+ free and don't need audio, 26B A4B is the stronger pick.
What's the best Gemma 4 model for coding?
31B leads on LiveCodeBench v6 (80.0%) and Codeforces Elo (2150), with 26B A4B close behind (77.1% / 1718). On smaller hardware, 12B (72.0% / 1659) is a credible coding assistant, and a community fine-tune (Gemma 4 12B Coder Fable5) sharpens it further specifically for Python.