Gemma 4 31B

Frontier intelligence in a 31 billion parameter model

Gemma 4 31B delivers exceptional performance across reasoning, coding, and multimodal tasks. With 256K context window and support for 140+ languages, it ranks #3 on Arena AI leaderboard while outperforming models 20x larger.

Model variants

Instruction-tuned and base models

Choose between the instruction-tuned variant optimized for chat and task completion, or the base model for fine-tuning and specialized applications.

Dense Architecture

30.7B parameters of pure reasoning power

Gemma 4 31B uses a dense architecture optimized for complex reasoning, coding, and multimodal understanding.

Best for production deployments requiring maximum intelligence and reliability.

Instruction-tuned

31B Instruct

Optimized for conversational AI and complex task completion

Fine-tuned with RLHF for following instructions and multi-turn dialogue

Available now

Pre-trained

31B Base

Foundation model for fine-tuning and specialized applications

Pre-trained on diverse multimodal data for maximum flexibility

Available now

Capabilities

Frontier-level performance across reasoning, coding, and multimodal tasks

Gemma 4 31B combines advanced reasoning, exceptional coding abilities, and multimodal understanding in an efficient architecture.

Advanced reasoning

Built-in thinking mode enables step-by-step reasoning. Achieves 89.2% on AIME 2026 mathematics benchmark.

Exceptional coding

80% on LiveCodeBench v6 and 2150 Codeforces ELO. Native function calling for agentic workflows.

Multimodal understanding

Processes text and images with variable aspect ratios. 76.9% on MMMU Pro multimodal reasoning.

256K context window

Extended context for long documents, codebases, and multi-turn conversations.

140+ languages

Multilingual support with cultural context understanding. 88.4% on MMMLU benchmark.

Efficient deployment

Optimized architecture with hybrid attention mechanism for fast inference and low memory footprint.

Key highlights

Exceptional performance metrics

Gemma 4 31B achieves frontier-level results across diverse benchmarks while maintaining efficient resource usage.

Top achievements

  • #3 on Arena AI leaderboard (ELO 1452)
  • 89.2% on AIME 2026 mathematics
  • 80% on LiveCodeBench v6 coding
  • 84.3% on GPQA Diamond scientific knowledge
  • 86.4% on τ2-bench agentic tool use

Technical specs

  • 30.7B parameters with dense architecture
  • 256K token context window
  • Support for 140+ languages
  • Hybrid attention mechanism
  • Variable image resolution support

Performance

Ranks #3 on Arena AI, outperforming models 20x larger

Gemma 4 31B achieves frontier-level performance across reasoning, coding, and multimodal tasks with exceptional efficiency.

Gemma 4 31B demonstrates consistent excellence across reasoning, coding, multimodal, and agentic benchmarks.

Gemma 4 31B performance comparison chart

Arena AI ELO 1452 - #3 ranked open model as of April 2, 2026

89.2% on AIME 2026 mathematics (no tools)

80% on LiveCodeBench v6 competitive coding

84.3% on GPQA Diamond scientific knowledge

86.4% on τ2-bench agentic tool use

Benchmark comparison

Comprehensive evaluation across key benchmarks

Gemma 4 31B demonstrates consistent excellence across reasoning, coding, multimodal, and agentic tasks.

Benchmark
Gemma 4 31B IT
Thinking
Featured
Gemma 4 26B A4B IT
Thinking
Gemma 4 E4B IT
Thinking
Gemma 3 27B IT
Arena AI (text)
As of April 2, 2026
14521441-1365
MMMLU
Multilingual Q&A
No tools
85.2%82.6%69.4%67.6%
MMMU Pro
Multimodal reasoning
76.9%73.8%52.6%49.7%
AIME 2026
Mathematics
No tools
89.2%88.3%42.5%20.8%
LiveCodeBench v6
Competitive coding
80.0%77.1%52.0%29.1%
GPQA Diamond
Scientific knowledge
No tools
84.3%82.3%58.6%42.4%
τ2-bench
Agentic tool use
Retail
86.4%85.5%57.5%6.6%

Benchmark results from official Gemma 4 model card. Arena AI scores as of April 2, 2026.

Advanced Reasoning

Step-by-step thinking for complex problems

Gemma 4 31B features configurable thinking modes that enable transparent reasoning processes for mathematics, logic, and multi-step problem solving.

  • 89.2% accuracy on AIME 2026 mathematics benchmark
  • Built-in reasoning mode with step-by-step explanations
  • Excels at scientific knowledge and logical deduction
Step-by-step thinking for complex problems

Coding Excellence

Elite performance on competitive programming

With 80% on LiveCodeBench v6 and 2150 Codeforces ELO, Gemma 4 31B excels at code generation, debugging, and agentic workflows with native function calling.

  • 80% on LiveCodeBench v6 competitive coding problems
  • 2150 Codeforces ELO rating
  • Native function calling for autonomous agents
Elite performance on competitive programming

Multimodal Understanding

Text and image processing with variable resolution

Process text and images together with support for variable aspect ratios and resolutions. Excels at document parsing, OCR, and visual reasoning.

  • 76.9% on MMMU Pro multimodal reasoning
  • Variable image resolution support (70-1120 tokens)
  • Document parsing, OCR, and chart comprehension
Text and image processing with variable resolution

Download weights

Self-hosted deployment

Download official model weights for deployment on your infrastructure.

Deploy and scale

Production deployment options

Enterprise-ready deployment on Google Cloud, Kubernetes, or your own infrastructure.

Join the Gemmaverse

Part of the broader Gemma ecosystem

Gemma 4 31B is part of Google's open model family, with extensive community support, integrations, and resources.

Documentation

Complete guides for integration and deployment

Read docs

Safety & Responsibility

Ethical AI development and safety guidelines

Learn more

Model Card

Technical specifications and evaluation results

View details

Google AI Studio

Try Gemma models with no setup

Try now

Community

Join developers building with Gemma

Explore

Cloud Deployment

Enterprise deployment on Google Cloud

Deploy

Get started

Ready to build with Gemma 4 31B?

Start chatting with Gemma 4 31B now, or download the model for self-hosted deployment on your infrastructure.