Key takeaways

• Gemma 4 31B is a 30.7B dense multimodal Transformer-based model by Google DeepMind, designed for text and image input with text output, focusing on coding, reasoning, and document understanding across 140+ languages. It features a 256K token context window, configurable reasoning modes, native function calling, and open weights under the Apache 2.0 license.
• The model demonstrates exceptional mathematical problem-solving capabilities, achieving 96.6% accuracy on the MATH-500 benchmark, particularly strong in Algebra, Prealgebra, and Number Theory with robust step-by-step reasoning. Its strengths include excellent performance across all Algebra levels and strong problem-solving methodology with clear, step-by-step explanations.
• A significant limitation is its performance in the Humanity's Last Exam benchmark, scoring only 10.33% accuracy. Common failure modes include numerical calculation errors (even when the method is correct), precise factual recall issues, difficulties interpreting complex, nuanced, and multimodal problem statements, and misinterpreting problem constraints or diagrams in geometry. The model's theoretical understanding often doesn't translate to correct final answers due to these issues.
• The model shows strong reasoning and correctly identifies relevant scientific/mathematical principles in many cases on the Humanity's Last Exam, but these are often undermined by consistent errors in final arithmetic calculations. It struggles with precise factual recall for specific values or names in specialized academic domains and misinterprets image-based questions.
• Overall, Gemma 4 31B excels in structured mathematical reasoning but requires improvement in meticulous arithmetic for complex expressions, precise interpretation of geometric setups and visual cues, and fine-grained contextual understanding, especially in multimodal and highly nuanced academic problems.

Technical specifications

Context length

262k tokens

Max output

—

Architecture

Transformer

Modalities

text+image+video->text

Parameters

~31B

Open weights

Yes

Licence

Apache 2.0

Regions

USA

Provider

Google

Provider model key

google/gemma-4-31b-it