Wire flash
TechOrcaRouter open-sources uncensored GLM-5.3-Flash, revealing Zhipu AI safety alignment embedded deep in model core
Editorial responsibility
- No named human review is recorded for this page.
- Source reporting is collected, normalized, translated or condensed automatically when needed.
- Automatically published source-backed update
A user or group called @OrcaRouter has open-sourced an 'uncensored' version of the GLM-5.3-Flash model, named GLM-5.3-Flash-Uncensored-FP8. The model has 320 billion total parameters with 18 billion activated, and was directly modified on the official native block-FP8 weights without using LoRA or jailbreak techniques. Notably, the developers discovered that Zhipu AI, the original creator of GLM-5.3-Flash, may have embedded safety alignment deeper into the model than previously expected, suggesting that removing safety constraints is more complex than simple surface-level modifications. This release highlights ongoing tensions in the AI community between open access and built-in safety measures.
Source report
A new "uncensored version" of GLM-5.3-Flash has been released, revealing interesting insights into Zhipu's safety alignment implementation.
Key Findings
- Researchers discovered that Zhipu may have embedded safety alignment deeper within the model than previously assumed.
- The uncensored version was achieved through direct modification of the official native block-FP8 weights, without using LoRA or jailbreak techniques.
Technical Details
- Model: GLM-5.3-Flash-Uncensored-FP8
- Open-sourced by: @OrcaRouter
- Total parameters: 320B
- Activated parameters: 18B
Source
MaxForAINeutral / independent
Part of this Story
Zhipu AI Open-Sources GLM-5.3-Flash, a Cost-Efficient Multimodal Model