Google Gemma 4 31B Instruct
Google’s open-weights dense multimodal model with reasoning, tool use, and 256K context.
Overview
What is Google Gemma 4 31B Instruct
Google Gemma 4 31B Instruct is an open-weights dense language model built by Google DeepMind. Released in April 2026 under Apache 2.0, it handles text, image, and video input, supports reasoning, tool use, and web search, and is designed for coding, agentic workflows, and multilingual tasks across 140+ languages.
Running it privately on Venice
On Venice, Gemma 4 31B runs privately with zero retention — your prompts are not stored or used for training. You get the full open-weights model with vision, reasoning, function calling, and web search capabilities at a fraction of the cost of closed rivals, with no account-linked history.
Assessment
Strengths and limitations
- Fully open weights under Apache 2.0: auditable, fine-tunable, and deployable anywhere without vendor lock-in.
- Multimodal reasoning with vision, video input, multiple image inputs, and native function calling for agentic workflows.
- Dense 31B architecture delivers consistent latency and simpler deployment than MoE counterparts.
- Configurable reasoning modes let you balance depth of thought against speed.
- Supports structured JSON output, web search, and log probabilities for building autonomous pipelines.
- Strong multilingual coverage across 140+ languages.
- Dense 31B size demands more VRAM than MoE alternatives with similar active parameters, making local hosting costlier.
- Not fully uncensored: base model safety filters remain active.
- Audio input is not natively supported on the 31B variant (only on smaller Gemma 4 models).
- Self-hosting at full bf16 precision requires high-end GPU infrastructure.
Capabilities
What it supports
- Tool use / function calling
- Vision (image input)
- Reasoning
- Web search
- Code-optimized
- Structured output (JSON schema)
- Audio input
- Video input
- Multiple image inputs
- Log probabilities
Specifications
Datasheet
- Maker
- Google DeepMind
- Released
- April 2, 2026
- Architecture
- Dense transformer
- Parameters
- 31 billion
- License
- Apache 2.0
- Open weights
- Yes
- Modality
- Text, image, and video input; text output
- Context window
- 256K tokens
- Max output
- 8.192K tokens
- Capabilities
- Vision, Function calling, Reasoning, Web search
- Privacy on Venice
- Private — zero retention
- Available on Venice since
- Apr 2026
API
Call it from your code
Venice exposes an OpenAI-compatible API. Point your base URL at Venice and pass the model id.
curl https://api.venice.ai/api/v1/chat/completions \
-H "Authorization: Bearer $VENICE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "google-gemma-4-31b-it",
"messages": [{ "role": "user", "content": "Explain quantum tunneling simply." }]
}'Pricing
What it costs on Venice
Billed per token on Venice: $0.12 per 1M input tokens and $0.36 per 1M output tokens.
New Venice accounts include a free daily allowance and 500 welcome credits — no credit card required.
Alternatives
How it compares
| Model | Context window | Open weights | Best for | Price (Venice) |
|---|---|---|---|---|
| Gemma 4 31B Instruct | 256K tokens | Yes | Open multimodal reasoning | $0.12 in · $0.36 out / 1M |
| DeepSeek V3.2 | 160K tokens | Yes | Efficient MoE coding | $0.33 in · $0.48 out / 1M |
| Kimi K2.6 | 256K tokens | Yes | Long-context agent tasks | $0.75 in · $3.50 out / 1M |
| Claude Sonnet 4.6 | 1M tokens | No | Proprietary reasoning | $3.60 in · $18 out / 1M |
The best open-weights choice on Venice for vision, tool use, and web search at low cost.
Use cases
What it is good for
- 01Coding assistance and autonomous agent pipelines with function calling and structured output.
- 02Multilingual content generation and long-document analysis across 140+ languages.
- 03Vision-based workflows — interpreting charts, diagrams, and video alongside text.
- 04On-prem or private-cloud deployment where open weights and Apache 2.0 licensing matter.
- 05Web-grounded research and reasoning tasks that benefit from configurable thinking modes.
Prompting
Getting better results
Use native system prompt support to set persistent behavior and output formats.
For coding or logic tasks, explicitly ask for step-by-step reasoning to engage thinking modes.
Leverage JSON schema mode when you need structured, parseable outputs.
Upload images at native resolution; the model handles variable aspect ratios automatically.
FAQ
Frequently asked questions
Google Gemma 4 31B Instruct is an open-weights dense language model built by Google DeepMind and released in April 2026 under Apache 2.0. It handles text, image, and video input, supports reasoning, tool use, and web search, and is designed for coding, agentic workflows, and multilingual tasks.
Venice bills per token with zero subscription required: $0.12 per 1M input tokens, $0.36 per 1M output tokens, and $0.09 per 1M cached input tokens.
Yes. It is released under Apache 2.0 with open weights, so you can download, self-host, fine-tune, and audit the model without vendor lock-in.
Yes. On Venice it supports function calling, vision (image and video input), multiple image inputs, structured JSON output, web search, reasoning, and log probabilities.
The model supports a context window of 256K tokens, with a maximum output of 8.192K tokens per generation on Venice.
No. Venice does not add extra censorship, but the model’s base safety filters remain active, so it is not fully uncensored.
Choose Gemma 4 31B if you need open-weights multimodal reasoning, vision, web search, and lower per-token pricing. Choose DeepSeek V3.2 if you prefer a MoE architecture optimized for coding throughput and can accept a higher per-token cost.
Yes. Venice runs it under a private zero-retention tier — your prompts are not stored, profiled, or used for training, and there is no account-linked history.
Run Google Gemma 4 31B Instruct privately
No prompt logging. No data used for training.