LLMPrivate

Google Gemma 4 31B Instruct

Google’s open-weights dense multimodal model with reasoning, tool use, and 256K context.

Maker
Google DeepMind
Modality
Text + vision
Context
256K tokens
Open weights
Yes

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.

Private (zero retention)No prompt trainingTEE · hardware enclaveEnd-to-end encrypted

Assessment

Strengths and limitations

Strengths
  • 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.
Limitations
  • 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.

Input / 1M tokens
$0.12
Per 1M tokens
Output / 1M tokens
$0.36
Per 1M tokens
Cached input / 1M
$0.09
Per 1M tokens

New Venice accounts include a free daily allowance and 500 welcome credits — no credit card required.

Alternatives

How it compares

ModelContext windowOpen weightsBest forPrice (Venice)
Gemma 4 31B Instruct256K tokensYesOpen multimodal reasoning$0.12 in · $0.36 out / 1M
DeepSeek V3.2160K tokensYesEfficient MoE coding$0.33 in · $0.48 out / 1M
Kimi K2.6256K tokensYesLong-context agent tasks$0.75 in · $3.50 out / 1M
Claude Sonnet 4.61M tokensNoProprietary 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

  1. 01Coding assistance and autonomous agent pipelines with function calling and structured output.
  2. 02Multilingual content generation and long-document analysis across 140+ languages.
  3. 03Vision-based workflows — interpreting charts, diagrams, and video alongside text.
  4. 04On-prem or private-cloud deployment where open weights and Apache 2.0 licensing matter.
  5. 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.