Gemma 4 26B A4B Uncensored
An uncensored, privacy-first Mixture-of-Experts (MoE) model based on Google's Gemma 4, optimized for unbiased reasoning, coding, and search.
Overview
What is Gemma 4 26B A4B Uncensored
Gemma 4 26B A4B Uncensored is a community-modified, de-aligned version of Google's 26B Mixture-of-Experts model. Fine-tuned using the 'Heretic' ablation method, it bypasses standard safety filters to deliver uncensored, highly objective reasoning, coding, and multilingual text generation without corporate refusals.
Running it privately on Venice
On Venice, this model runs with absolute privacy inside a hardware-isolated Trusted Execution Environment (TEE) with end-to-end encryption. Because Venice enforces zero retention, your prompts are never stored, logged, or used to train future models, ensuring complete sovereignty over your uncensored workflows.
Assessment
Strengths and limitations
- Uncensored 'Heretic' fine-tune drastically reduces refusals for unbiased, raw creative and analytical tasks.
- Highly efficient Mixture-of-Experts (MoE) architecture utilizing 8 active experts out of 128 total.
- Hardware-enforced privacy running in TDX-attested TEE enclaves with end-to-end encryption.
- Integrated Web Search capability for real-time, grounded information retrieval.
- May generate highly controversial, graphic, or unsafe content due to the removal of alignment guardrails.
- Context window on Venice is limited to 64K tokens, compared to the base model's native 256K.
- Lacks native image or multimodal input support on this specific Venice text pipeline.
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 (Base) / Phala (Fine-tune)
- Released
- May 23, 2026
- Architecture
- Sparse Mixture-of-Experts (MoE)
- Parameters
- 25.2B total (3.8B active)
- Open weights
- No (Hosted pipeline)
- Context window
- 64K tokens
- Max output
- 4.096K tokens
- Capabilities
- 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-4-26b-a4b-uncensored-p",
"messages": [{ "role": "user", "content": "Explain quantum tunneling simply." }]
}'Pricing
What it costs on Venice
Billed per token on Venice: $0.19 per 1M input tokens and $0.88 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 | Strongest at | Open weights | Price (Venice) |
|---|---|---|---|---|
| Gemma 4 26B A4B Uncensored | 64K tokens | Uncensored reasoning & TEE privacy | No | $0.19 in · $0.88 out / 1M |
| Google Gemma 4 31B Instruct | 256K tokens | Aligned reasoning & coding | Yes | $0.12 in · $0.36 out / 1M |
| DeepSeek V3.2 | 160K tokens | Coding & math | Yes | $0.33 in · $0.48 out / 1M |
| GLM 5.1 | 200K tokens | Multilingual reasoning | Yes | $1.10 in · $4.15 out / 1M |
The uncensored, MoE-powered privacy champion.
Use cases
What it is good for
- 01Uncensored creative writing, roleplay, and brainstorming free from corporate safety filters.
- 02Objective analysis of controversial historical, political, or philosophical topics.
- 03Coding assistance and debugging where standard models refuse due to security-related keywords.
- 04Real-time research and fact-checking using integrated web search.
Prompting
Getting better results
Use a system prompt to define the exact persona; this model natively supports system prompts across 35+ languages.
Be direct and explicit—no need to use euphemisms or bypass phrasing since the model is uncensored.
Enable web search when asking about events after its April/May 2026 cutoff.
Version history
Official base MoE model released by Google DeepMind.
Community de-aligned variant fine-tuned by Phala.
FAQ
Frequently asked questions
Gemma 4 26B A4B Uncensored is a community-modified, de-aligned version of Google's 26B Mixture-of-Experts (MoE) model. Fine-tuned using the 'Heretic' ablation method, it bypasses standard safety filters to deliver uncensored, highly objective reasoning, coding, and multilingual text generation without corporate refusals.
On Venice, the model is billed per token at $0.19 per 1M input tokens and $0.88 per 1M output tokens, making it highly cost-effective.
The base model is open-weights (Apache 2.0), but this specific hosted pipeline on Venice is served via a third-party provider. You can try it on Venice using free daily credits or a premium subscription.
It means the model has undergone Arbitrary-Rank Ablation (ARA) to remove safety alignment guardrails. It will answer sensitive, controversial, or complex prompts that standard Google models refuse.
The 31B Instruct is Google's official, aligned dense model with a 256K context. The 26B Uncensored is a community MoE variant with a 64K context on Venice that does not refuse prompts.
Yes, this model supports integrated web search on Venice, allowing it to fetch real-time information to ground its answers.
It runs in a hardware-isolated Trusted Execution Environment (TEE) with end-to-end encryption. Venice maintains a zero-retention policy, meaning your prompts are never stored or used for training.
Run Gemma 4 26B A4B Uncensored privately
No prompt logging. No data used for training.