🧠

Multimodal Language Models - Gemma 4 31B

🛠

Advanced Text, Image & Video Understanding with Built-In Reasoning

Google DeepMind's most capable open model, a multimodal AI that reads, reasons, codes, and understands images and video across reasoning, coding, and vision benchmarks.

What it does

Gemma 4 31B is Google DeepMind's flagship open-weight AI model, the most powerful in the Gemma 4 family. It is a fully multimodal, instruction-tuned model that understands and generates text, analyzes images and video, reads documents, writes and debugs code, performs complex multi-step reasoning, and can operate as an autonomous AI agent using built-in tool-calling capabilities.

Problem it solves

  • Frontier-level reasoning in an open model – Achieves 89.2% on AIME 2026 (advanced mathematics) and 84.3% on GPQA Diamond (expert science questions), putting it in rare company among open-source models and competitive with leading closed-source alternatives
  • Deep multimodal understanding – Processes text, images, and video together in a single conversation, enabling tasks that require understanding across multiple types of content simultaneously
  • Very long document handling – A 256K token context window means it can read and reason over extremely long documents, reports, codebases, or conversation histories in a single pass
  • Coding and software development – Achieves a Codeforces ELO of 2150, placing it at an elite competitive programming level — suitable for real-world code generation, debugging, and review
  • Agentic workflows – Native function-calling support means it can be connected to external tools, APIs, and data sources to act autonomously on complex, multi-step tasks
  • Multilingual reach – Out-of-the-box support for 35+ languages, pre-trained on 140+ languages, making it suitable for global deployments without additional localization work

Input/Output

  • Input:
    • Text – Questions, instructions, documents, code, and conversations
    • Images – Photos, screenshots, charts, diagrams, documents, and UI screens at variable resolution
    • Video – Video content analyzed as sequential frames
  • Advanced:
    • Max New Tokens: Controls the maximum length of the model's response. For most tasks, 32,768 tokens is recommended. For complex math or coding competition problems, up to 81,920 tokens is advised to give the model sufficient space to reason thoroughly.
    • Temperature: Controls how creative or predictable the response is. Set to 0.6 for precise coding tasks where accuracy is critical. Use 1.0 for general tasks and open-ended reasoning where more varied responses are beneficial.
    • Top-P (Nucleus Sampling): Sets a probability threshold that limits which words the model considers for its next response. For example, a Top-P of 0.95 means the model only picks from the most likely words that together add up to 95% probability, filtering out unlikely or off-topic choices. Lower values make responses more focused and predictable; higher values allow more variety. Recommended: 0.95 for thinking mode, 0.8 for standard instruct mode.
    • Top-K Sampling: Limits the model to choosing only from the top K most likely next words at each step. For example, Top-K of 20 means the model only considers the 20 most probable options, ignoring everything else. Works alongside Top-P to keep responses coherent and on-topic. Recommended: 20 across all modes.
    • Enable Thinking: When turned on, the model reasons through the problem step-by-step before delivering its final answer. This significantly improves accuracy on complex tasks like coding, mathematics, and multi-step reasoning. When turned off, the model responds directly and more quickly, which is better suited for simple, conversational, or time-sensitive tasks. On by default in Qwen3.6.
  • Output:
    • Generated text responses, analysis, summaries, code, or structured data
    • Supports a built-in thinking mode where the model reasons step-by-step before delivering its final answer, improving accuracy on complex tasks

Accuracy & Speed

Benchmark Performance

Benchmark
Gemma 4 31B
What it measures
MMLU Pro
85.2%
Broad expert knowledge
AIME 2026
89.2%
Advanced mathematics
LiveCodeBench v6
80.0%
Real-world coding ability
Codeforces ELO
2,150
Competitive programming
GPQA Diamond
84.3%
Expert-level science reasoning
MMMU Pro
76.9%
Multimodal understanding
MMMLU (multilingual)
88.4%
Cross-language knowledge
BigBench Extra Hard
74.4%
Complex reasoning tasks

Technical Details

Model Size
30.7 billion parameters
Context window
256,000 tokens (~200,000 words)
Modalities supported
Text, Image, Video
Thinking mode
Built-in — enable for step-by-step reasoning on complex tasks
Function calling
Native — connects to external tools and APIs for agentic workflows
Languages
35+ out-of-the-box, pre-trained on 140+ languages
Supported frameworks
Hugging Face Transformers, vLLM, and others

Compliance & Provenance

Provider
Google (open weights)
Provider type
GPAI — General-Purpose AI (GPAI)
License
Apache 2.0
EU AI Act risk class
Limited Risk
Art. 50 transparency
Required — outputs are marked. See AI Policy §2.
Region availability
Available globally
Training data summary

For more on how we classify models and mark outputs, see our AI Policy.

Model Source