Gemma 4 31B Instruct
Google DeepMind's dense 31B-parameter open-source instruct model with strong reasoning and web search capabilities.
Overview
What is Gemma 4 31B Instruct
Gemma 4 31B Instruct is a 31-billion-parameter dense language model from Google DeepMind, released in April 2026 under the Apache 2.0 open-source license. It supports advanced reasoning and web search on Venice, offering a transparent, cost-efficient alternative to closed frontier models while running with zero prompt retention.
Running it privately on Venice
On Venice, Gemma 4 31B Instruct runs inside a trusted execution environment with end-to-end encryption and zero retention — your prompts are never stored or used for training. You get the same open-source weights Google DeepMind published, executed privately without Big-Tech surveillance. It is an ideal choice for builders who value sovereignty and permissionless access.
Assessment
Strengths and limitations
- Open-source Apache 2.0 weights let you inspect, fine-tune, or self-host the model outside Venice.
- Built-in reasoning mode and web search support on Venice for research-heavy and analytical workflows.
- Runs inside a TEE with end-to-end encryption on Venice — your prompts are never stored or used for training.
- Extremely cost-efficient token pricing compared to closed frontier alternatives.
- Venice hosts a 32K context window, which is shorter than some rivals offering 160K–1M tokens.
- Not uncensored: safety alignment is present, so it may refuse certain sensitive or harmful requests.
- Smaller 31B dense architecture lags behind the absolute top-tier closed models on the most complex reasoning benchmarks.
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 2026
- Architecture
- Dense
- Parameters
- 31B
- Open weights
- Yes — Apache 2.0
- Context window
- 32K tokens
- Max output
- 4.096K tokens
- Capabilities
- Reasoning, Web search
- Privacy on Venice
- Private — zero retention
- Available on Venice since
- May 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": "e2ee-gemma-4-31b",
"messages": [{ "role": "user", "content": "Explain quantum tunneling simply." }]
}'Pricing
What it costs on Venice
Billed per token on Venice: $0.14 per 1M input tokens and $0.43 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 | Price (Venice) | Best for |
|---|---|---|---|---|
| Gemma 4 31B Instruct | 32K tokens | Yes | $0.14 in · $0.43 out / 1M | Private open-source reasoning |
| DeepSeek V3.2 | 160K tokens | Yes | $0.33 in · $0.48 out / 1M | Long-context coding |
| Claude Sonnet 4.6 | 1M tokens | No | $3.60 in · $18 out / 1M | Enterprise reasoning |
| Kimi K2.6 | 256K tokens | Yes | $0.75 in · $3.50 out / 1M | Long-document analysis |
The only open-source model in this set that runs with zero retention and TEE privacy on Venice.
Use cases
What it is good for
- 01Long-form writing and content drafting with web-grounded facts.
- 02Open-source prototyping and fine-tuning for domain-specific chatbots.
- 03Privacy-sensitive research where zero retention and TEE isolation are required.
- 04Cost-efficient production text generation at high volume.
Prompting
Getting better results
Enable reasoning mode for math, logic, or multi-step coding problems.
Use the web search capability to ground answers in current information beyond the training cutoff.
Keep inputs within the 32K context limit; summarize long documents before sending.
Version history
Previous generation open-weights model.
Current generation with reasoning mode and web search support on Venice.
FAQ
Frequently asked questions
Gemma 4 31B Instruct is a 31-billion-parameter dense language model built by Google DeepMind. Released in April 2026 under Apache 2.0, it is designed for reasoning, coding, and general instruction-following, and is available on Venice with web search support.
Venice bills per token: $0.14 per 1M input tokens, $0.43 per 1M output tokens, and $0.03 per 1M cached input tokens. There is no subscription required.
The model is open-source under the Apache 2.0 license, so you can download the weights and self-host or fine-tune them. On Venice you pay only for the tokens you use; there is no free tier beyond any welcome credits Venice offers.
On Venice, Gemma 4 31B Instruct currently supports reasoning and web search, but not tool use or vision. If you need native tool calling, look for models that list it in their Venice capability metadata.
It runs under Venice’s private tier with zero retention, inside a trusted execution environment (TEE) with end-to-end encryption. Your prompts are not stored, profiled, or used to train future models.
Claude Sonnet 4.6 generally leads on the hardest reasoning tasks and offers a 1M-token context window, but it is closed-source and far more expensive. Gemma 4 31B Instruct wins on cost, open weights, and privacy, making it ideal for builders who need a transparent, self-hostable model.
Venice currently hosts Gemma 4 31B Instruct with a 32K-token context window and a maximum output of 4,096 tokens per generation.
Run Gemma 4 31B Instruct privately
No prompt logging. No data used for training.