Aoxo Breaks Barriers With Sarvam-30b-Uncensored AI Weights

The Sarvam-30b-uncensored project delivers a modified version of the original Sarvam-30B system, specifically engineered to operate without standard safety filters. By removing internal alignment constraints, the architecture outputs direct responses across all prompts while keeping its reasoning structure completely intact.

Released by independent developer Aoxo, this update addresses the demand for unrestricted models that maintain high-quality generation. The modification targets local workflows that require complete control over text processing without relying on external moderation layers.

Model Size: ~60GB & VRAM GPU: requirements vary

Direct output without refusal mechanisms

• Preserves multilingual functionality across twenty-two regional languages.
• Retains full coding and autonomous task capabilities.
• Maintains a sixty-five thousand token context window for long documents.
• Uses precise weight adjustments to bypass standard alignment layers.
• Supports custom instruction tuning without pre-programmed compliance rules.

Professionals managing private databases will find this setup valuable for tasks that require uninterrupted text generation. Local workstation operators can run the model off the grid to handle sensitive prompts without depending on cloud-based safety checks.

Technical notes from the creator

The development process uncovered two separate refusal pathways inside the original architecture. One pathway shapes the internal reasoning steps, while another controls the final output. The developer stated in a post that

"reasoning models have 2 refusal circuits, not one."

Running the adjusted system requires users to build their own moderation tools if they plan to share the results publicly. The stripped weights lack built-in guardrails, which makes careful testing essential before deployment.

You can access the adjusted weights directly through the Hugging Face repository.