LLMPrivate

Gemma 4 Uncensored

A community derivative of Google's Gemma 4 26B MoE with reduced safety alignment, offering 256K context, vision, and tool use at very low cost.

Maker
Google DeepMind (community derivative)
Modality
Text + vision
Context
256K tokens
Open weights
Yes — Apache 2.0

Overview

What is Gemma 4 Uncensored

Gemma 4 Uncensored is a community derivative of Google DeepMind's Gemma 4 26B MoE, modified to reduce refusals and safety alignment. It retains the family's 256K context, vision, function calling, and multilingual support, while running as open weights under an Apache 2.0 license.

Running it privately on Venice

On Venice, Gemma 4 Uncensored runs under a private, zero-retention privacy tier — your prompts are not stored or used for training. You get vision, tool use, and web search capabilities at $0.16 per 1M input tokens, with no Big-Tech surveillance. It is served as open-weights int4 quantization, so you can use it permissionlessly without an account lock-in.

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

Assessment

Strengths and limitations

Strengths
  • Open-weights Apache 2.0 license enables permissionless local and cloud deployment.
  • Extremely low inference cost on Venice with 256K context and vision support.
  • Native tool use, function calling, web search, and structured JSON output for agentic workflows.
  • Strong reasoning and multimodal capabilities with image input and support for 140+ languages.
  • Reduced refusal behavior compared to the official instruct variant, giving more direct answers.
  • Efficient MoE architecture: 26B total with only ~4B active parameters per token.
Limitations
  • Community derivative, not an official Google release — safety alignment has been intentionally weakened, so outputs require user discretion.
  • Quantized to int4 on Venice, which may trade a small amount of precision for speed and VRAM efficiency.
  • Active parameter count (~4B) is lower than the dense 31B variant, potentially capping peak reasoning on the hardest tasks.
  • Privacy is private/zero-retention but does not run inside a TEE or end-to-end encryption layer on Venice.
  • As with all modified derivatives, factual accuracy and hallucination risks remain unchanged from the base model.

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 (community derivative)
Released
April 2, 2026 (Gemma 4 family)
Architecture
Transformer, Mixture-of-Experts (MoE)
Parameters
26B total (~4B active per token)
Modality
Text, image input (vision)
Open weights
Yes — Apache 2.0
Quantization on Venice
int4
Context window
256K tokens
Max output
8.192K tokens
Capabilities
Vision, Function calling, 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": "gemma-4-uncensored",
    "messages": [{ "role": "user", "content": "Explain quantum tunneling simply." }]
  }'

Pricing

What it costs on Venice

Billed per token on Venice: $0.16 per 1M input tokens and $0.50 per 1M output tokens.

Input / 1M tokens
$0.16
Per 1M tokens
Output / 1M tokens
$0.50
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 weightsPrice (Venice)Best for
Gemma 4 Uncensored256K tokensYes$0.16 in · $0.50 out / 1MLow-cost uncensored derivative
Google Gemma 4 31B Instruct256K tokensYes$0.12 in · $0.36 out / 1MOfficial dense variant
DeepSeek V3.2160K tokensYes$0.33 in · $0.48 out / 1MGeneral open-weight inference
Claude Sonnet 4.61M tokensNo$3.60 in · $18 out / 1MLong-context closed model

The budget-friendly open-weights choice with reduced refusals and full multimodal support.

Use cases

What it is good for

  1. 01Local or private cloud AI agents that need tool calling and vision without vendor lock-in.
  2. 02Long-document analysis and summarization across 256K tokens.
  3. 03Multilingual content generation and translation at very low cost.
  4. 04Coding assistants where reduced refusal behavior speeds up iterative debugging.
  5. 05Research and creative writing workflows that benefit from direct, unfiltered responses.

Prompting

Getting better results

Be direct and specific; the model is tuned to skip excessive disclaimers and padding.

For coding tasks, ask for step-by-step reasoning before the final code block to improve accuracy.

Use the native JSON schema / structured output mode when building agents.

Leverage vision by attaching images and asking for structured extraction or description.

Take advantage of the 256K context by pasting long documents or multiple files in a single prompt.

Version history

Gemma 4 26B IT (official)
2026-04

Official Google release with standard safety alignment.

Gemma 4 Uncensored (derivative)
2026-04

Community derivative with reduced refusals and alignment behavior.

FAQ

Frequently asked questions

Gemma 4 Uncensored is a community derivative of Google DeepMind's Gemma 4 26B MoE, modified to reduce safety refusals and alignment behavior. It retains the base model's 256K context, vision, tool use, and Apache 2.0 open-weights license.

Venice bills per token at $0.16 per 1M input tokens and $0.50 per 1M output tokens. There is no subscription required; you pay only for what you use.

The base Gemma 4 weights are released under Apache 2.0, and this derivative is distributed as open weights. You can download and run it locally or via the Venice API without proprietary restrictions.

It is a derivative with reduced safety fine-tuning and fewer refusals than the official instruct variant. However, it is still subject to the base model's knowledge limits and can hallucinate or produce inaccurate information.

Yes. On Venice it supports function calling, tool use, vision (image input), web search, structured JSON output, and log probabilities.

Choose the Uncensored derivative if you want fewer refusals and lower cost for agentic workflows. Choose the official 31B Instruct if you prefer Google's dense architecture with standard safety alignment and slightly lower pricing.

Venice runs it under a private, zero-retention tier — your prompts are not stored, profiled, or used for training. Note that this deployment does not use a TEE or end-to-end encryption.

It supports up to 256K tokens of context, with a maximum output of 8,192 tokens per generation.

Run Gemma 4 Uncensored privately

No prompt logging. No data used for training.