Gemma 4 31B
Google · Text generation · 31B · 262k context · Released 30 July 2026
Gemma 4 31B is the largest dense model in the Gemma 4 release (the family also includes the smaller E2B and E4B, a 12B, and a 26B-A4B mixture-of-experts). It takes text and images and produces text, with a 256k context. For local use its appeal is twofold: it fits a single 24GB card at a 4-bit quantisation, and it is widely preferred in the community for creative writing and prose, though that is a matter of taste rather than a settled benchmark result. The licence is the other headline: Gemma 4 is Apache 2.0, dropping the custom Gemma Terms of Use earlier generations carried, subject to a separate Prohibited Use Policy.
Strengths
- Apache 2.0, a genuinely permissive licence and a change from older Gemma terms
- A community favourite for creative writing and prose
- Reads images as well as text, with a 256k context
- Fits a single 24GB card at a 4-bit quantisation
Weaknesses
- A dense 31B, so no mixture-of-experts speed benefit for its size
- Its creative-writing standing is community preference, not a settled benchmark
- The Apache grant carries a separate Gemma Prohibited Use Policy
- Image support in local runtimes is less mature than text-only use
Hardware requirements
| Quantisation | Approx. VRAM | Notes |
|---|---|---|
| Q4_K_M | ~18GB | Common default; fits a 24GB card with headroom. Approximate |
| Q8_0 | ~33GB | Near-lossless, needs a 48GB card or more. Approximate |
| BF16 | ~62GB | Full precision, multi-GPU or server territory. Approximate |
Optimised builds available for Apple Silicon.
What you'd need to run this
Roughly what a machine to run this would need, at up to three levels of quality. Memory is the deciding factor.
Minimum to run it
Q4_K_M · ~18GB needed
One 24GB GPU
NVIDIA Tesla P40or a Mac or mini-PC with unified memory, if you prefer no discrete GPU, Mac mini M5 Pro .
32–64GB of system RAM alongside the card.
around £700–£1,100
What else 24GB runs →For good quality
Q8_0 · ~33GB needed
64GB of unified memory
Mac mini M4 Proor a single high-memory workstation card, NVIDIA RTX 6000 Ada Generation .
Unified memory is shared with the model, so it is already counted above.
Best quality
BF16 · ~62GB needed
96GB of unified memory
AMD Ryzen AI Max+ 395 (Strix Halo)or an 80GB-class data-centre card, which is usually rented by the hour, NVIDIA A100 80GB .
Unified memory is shared with the model, so it is already counted above.
Licence
Apache 2.0 — read the licence
Benchmarks
| Benchmark | Score | Source | As of |
|---|---|---|---|
| MMLU-Pro (thinking) | 85.2 | Gemma 4 technical report (Google) | July 2026 |
| GPQA Diamond (thinking) | 84.3 | Gemma 4 technical report (Google) | July 2026 |
| AIME 2026 (no tools) | 89.2 | Gemma 4 technical report (Google) | July 2026 |
| LMArena Elo (top dense open model) | 1451 | Gemma 4 technical report, citing LMArena | July 2026 |
Gemma 4 31B: common questions
- What hardware do I need to run Gemma 4 31B?
- At its most compressed (Q4_K_M) it needs roughly 18GB of VRAM, and about 24GB for good quality. VRAM figures are approximate and depend on context length and settings.
- Is Gemma 4 31B free for commercial use?
- Yes. Gemma 4 31B is licensed under Apache 2.0, which permits commercial use with no meaningful conditions.
- Can I run Gemma 4 31B on Apple Silicon?
- Yes. Gemma 4 31B has builds optimised for Apple Silicon, through MLX or GGUF on a Mac.
- What is Gemma 4 31B's context window?
- Gemma 4 31B has a context window of 262,144 tokens, about 262k.
Availability
Recommended for
- Local creative writing and prose on a 24GB card
- A permissively licensed general and image-reading model
- Long-context work that fits a single card
Related models
Related guides
Glossary
Catalogue entry last verified 2 October 2026. Specifications change; verify anything you are about to spend money on.