Unshackled Intelligence: The Abliterated Super-Series Release

The refusal layer is gone.

Jun Song has announced the release of a new tier of uncensored weights on Hugging Face. The goal is simple: remove the synthetic morality filters that throttle raw model capability.

The Payload

Three primary architectures have been processed through the abliteration pipeline:

• SuperGemma4-12b-abliterated: A lean, aggressive powerhouse optimized for zero-refusal logic.
• SuperMiniMax-M3-abliterated: High-capacity multimodal processing with extended context windows, now stripped of behavioral guardrails.
• SuperGLM-5.2-abliterated: A multimodal titan designed for complex reasoning and long-form synthesis, fully unconstrained.

The Process: Abliteration

Unlike traditional fine-tuning or RLHF, which attempt to "teach" a model to be helpful, abliteration targets the specific refusal vectors within the model s latent space. By neutralizing the directions that trigger "I cannot fulfill this request" responses, we recover the model s original, raw utility without degrading its core intelligence.

Compliance & Boundaries

While the Super-Series pushes the envelope, a hard line remains: Cybersecurity dataset versions will not be released. Due to current regulatory pressures and the risk of weaponization, these specific weights remain offline. The focus is on general intelligence, not exploit generation.

Infrastructure: c0mputeAI

For those lacking the local VRAM to run these beasts, c0mputeAI provides the infrastructure. Decentralized inference ensures that these models remain accessible and resistant to centralized kill-switches.

Status: Live on Hugging Face. Weights unlocked.