Model Details
Gemma 4 31B
Quick answer: Gemma 4 31B is the flagship dense open-weight model in Google DeepMind's Gemma 4 family and the successor to Gemma 3 27B. With 31B parameters, it delivers Gemma 4-generation training improvements — better coding, stronger multilingual capability, and improved instruction following — at the cost of a single 2× A100 or H100 80GB setup.
What Gemma 4 31B Is
Gemma 4 31B is Google DeepMind's largest dense open-weight model in the Gemma 4 generation. Building on the Gemma 3 27B architecture, the slight parameter increase (27B → 31B) reflects architecture refinements for better utilisation efficiency, while the training improvements from Gemma 4's development cycle deliver meaningful quality gains on reasoning, coding, and multilingual tasks.
For Benchgen evaluation purposes, Gemma 4 31B is the primary open-weight model to benchmark against the 30–35B closed-API tier (GPT-5.1 Instant, Claude Haiku 4.5).
Specifications
| Field | Value |
|---|---|
| Organization | Google DeepMind |
| Parameters | 31 billion |
| License | Gemma Terms of Use |
| Modality | Multimodal (text and vision) |
| Min VRAM (FP16) | ~62GB (2× A100 40GB or 1× H100 80GB) |
| HuggingFace | google/gemma-4-31b-it |
Related Models & Benchmarks
- Compare: Gemma 3 27B · Gemma 4 26B-A4B · Gemma 4 12B
- Benchmarks: MMLU-Pro · SWE-bench Verified
Last updated 2026-06-19.
Pricing
Open weights — self-host at no cost.
Gemma 4 31B vs Alternatives
| Model | MMLU-Pro | License |
|---|---|---|
| Gemma 4 31B | 85.2% | Gemma ToU |
| Gemma 4 26B A4B | — | — |
Check the MMLU-Pro leaderboard for current comparisons.
Frequently Asked Questions
What is Gemma 4 31B?
Google's Oct 2025 model scoring 85.2% MMLU-Pro. Gemma ToU license.Scores from Benchgen evaluations. Last updated 2026-07-24.