Google's Best Open Model Yet Has a Memory Problem

Google's Best Open Model Yet Has a Major Drawback

Google recently released the latest version of its open-source language model, Gemma-4-31B. This model is considered one of their best offerings in the open-weight space due to several notable features and performance improvements.

Key Features:

• Multimodal Capabilities: The 31B variant supports text, images, and video inputs.
• Large Context Window: It has a substantial KV cache footprint, allowing for extensive context handling.
• Built-in Thinking Mode: Users can enable or disable the thinking mode by including <|think|> in the system prompt. This feature outputs reasoning traces before providing final answers.

Performance Highlights:

• Coding Strengths: The model's Codeforces ELO of 2150 is a significant improvement over similar models at this size.
• Accuracy and Reasoning: It performs well on various benchmarks, including ranking above GLM-5 in certain evaluations.

Running the Model

To use Gemma-4-31B, you can load it through Hugging Face's standard Transformers interface:

python
 
1pip install -U transformers torch accelerate
2
3from transformers import AutoProcessor, AutoModelFor
4
5[Read the full article at AIModels.fyi](https://aimodels.substack.com/p/googles-best-open-model-yet-has-a)
6
7---
8
9**Want to create content about this topic?** [Use Nemati AI tools](https://nemati.ai) to generate articles, social posts, and more.