The newly launched Venice Uncensored 1.2, developed in collaboration with @dphnAI, features significant enhancements including upgraded vision support, a quadrupled context window, and improved tool-use capabilities. This model, which aims to provide the most uncensored version of Mistral 24B, has achieved 100% compliance on a comprehensive set of prompts, indicating its robust performance. The uncensoring was accomplished through a combination of supervised fine-tuning and KTO reinforcement learning, a method that avoids the common pitfalls associated with manual weight editing. Additionally, Venice Uncensored 1.2 was trained on Bittensor Subnet 4, representing a noteworthy application of decentralized infrastructure in AI development.

dphnAI: dphnAI is an AI lab focused on developing uncensored models and distributed inference technologies to enable user-controlled AI experiences. It partnered with Venice.ai to produce Venice Uncensored 1.2, applying supervised fine-tuning and KTO reinforcement learning for effective decensoring. The lab drives the Dolphin project, emphasizing steerable models without imposed ethics.
Dolphin: Dolphin is a project by dphnAI creating general-purpose, uncensored AI models that prioritize user control over system prompts and alignment. The Dolphin Mistral 24B Venice Edition forms the core of Venice Uncensored 1.2, fine-tuned to eliminate refusals while preserving base model capabilities. It promotes models as extensions of user intent, akin to tools like software or hardware.
Mistral: Mistral AI develops open-source large language models optimized for instruction-following and customization. Its Mistral 3.2 24B Instruct model serves as the foundation for Venice Uncensored 1.2, providing the core intelligence after fine-tuning removes safety refusals. The models support enterprise deployments and private tuning.
TargonCompute: TargonCompute is a secure cloud compute platform offering confidential GPU resources for AI workloads, integrated as Bittensor Subnet 4. It provided the compute infrastructure for training Venice Uncensored 1.2, supporting decentralized AI innovation. Recently, it released a whitepaper on decentralized compute using untrusted hardware.
Venice Uncensored: Venice Uncensored is an uncensored AI model serving as the default option on the Venice.ai platform, designed for private and unrestricted interactions supporting text, vision, and tools. Its 1.2 version, released today, was developed collaboratively with dphnAI and features upgrades like vision support, expanded context window, and stronger tool-use capabilities. Trained on Bittensor Subnet 4 via TargonCompute, it represents a step towards decentralized AI deployment.

`json
{
“Model Upgrades”: “Version 1.2 introduces vision support, a four times larger context window, and improved tool-use capabilities compared to previous versions.”,
“Uncensoring Approach”: “The model is uncensored using KTO reinforcement learning following supervised fine-tuning, avoiding direct weight edits with abliteration techniques which can cause bugs or performance degradation.”,
“Decentralized Infrastructure”: “Utilizes Bittensor Subnet 4 for training, emphasizing the use of decentralized resources in developing advanced AI models.”
}
`