# This Is What Happens When You CRUSH An AI Video Model **URL:** \ **Category:** Content Creators\ **Tags:** gpu, alex-ziskind, video, linux, quantisation\ **Created:** [18 June 2026 14:30 UTC](https://www.artofsm.art/t/this-is-what-happens-when-you-crush-an-ai-video-model/20353 "2026-06-18T14:30:34Z")\ **Posts on this page:** 2\ **Page:** 1
**Author:** ![artesia](https://www.artofsm.art/user_avatar/www.artofsm.art/artesia/32/36_2.png) [@artesia](https://www.artofsm.art/u/artesia)\ **Post date:** [18 June 2026 14:30 UTC](https://www.artofsm.art/t/this-is-what-happens-when-you-crush-an-ai-video-model/20353/1 "2026-06-18T14:30:34Z")
[![](https://www.artofsm.art/uploads/default/original/3X/0/d/0dc41e6c0e87c297405bd779571d84f64c2cccb0.jpeg "This Is What Happens When You CRUSH An AI Video Model") ](https://www.youtube.com/watch?v=octq1gJ0dkM) The video investigates how different quantization formats and bit depths affect the quality of AI video models, finding that integer-based quantization (Q8) outperforms floating-point (FP8) despite using the same bit count, and that audio quality degrades faster than video as quantization increases. It concludes that Q4 quantization offers the best balance between size and quality, especially for models with audio, and emphasizes the importance of choosing the right quantization format over simply reducing bit precision. ---
**Author:** ![artesia](https://www.artofsm.art/user_avatar/www.artofsm.art/artesia/32/36_2.png) [@artesia](https://www.artofsm.art/u/artesia)\ **Post date:** [18 June 2026 14:51 UTC](https://www.artofsm.art/t/this-is-what-happens-when-you-crush-an-ai-video-model/20353/3 "2026-06-18T14:51:13Z")
The video explores the impact of quantization on AI video models, specifically examining how reducing the bit precision of model weights affects video and audio quality. The creator tests two different video models—WAN 2.2 with 14 billion parameters and LTX 2.3 with 22 billion parameters—across an eight-step quantization ladder, from full precision (FP16) down to two bits. Both models are run on a powerful custom Linux setup with an RTX Pro 6000 GPU, ensuring that hardware limitations do not influence the results. The main variable changed is the quantization format and bit depth, allowing a clear comparison of quality degradation at each step. Initial tests at full precision (FP16 and BF16) show high-quality video and audio outputs, serving as baselines. When quantization is reduced to 8 bits, two formats are compared: FP8 (floating point 8-bit) and Q8 (integer 8-bit with group quantization). Surprisingly, Q8 outperforms FP8 in both video fidelity and audio accuracy, despite both using the same number of bits. FP8 even causes some bizarre errors, such as objects moving incorrectly in the video and audio artifacts like added laughter. This finding highlights that the quantization format is more critical than just the bit count. As the bit depth decreases further to Q6, Q5, and Q4, video quality gradually declines but remains usable, especially at Q4, which is popular in the community due to its balance between size and quality. The LTX model, which generates both video and audio, shows that audio quality degrades faster than video quality as quantization increases. This is significant because poor audio quality is more immediately noticeable and off-putting to viewers than slightly degraded video. Thus, for models supporting audio, maintaining higher audio fidelity is crucial. At the extreme low end of the quantization ladder—two bits per weight—the models become largely unusable. Video frames become inconsistent and distorted, character consistency is lost, and audio turns robotic or unintelligible. However, simple scenes with minimal motion, like a glass of water, still retain some semblance of quality even at two bits, indicating that the complexity of the scene influences how well low-bit quantization performs. The video also notes that some inherent model limitations, such as unrealistic physics in a marble rolling scene, are unrelated to quantization. The key takeaway from the experiment is that quantization format matters more than bit count, and that Q4 quantization strikes the best balance for most users, especially when audio is involved. Both tested models, despite their architectural differences, show similar degradation patterns, reinforcing the generality of these findings. The video encourages users to be mindful of audio quality when choosing quantization levels and to prefer integer-based quantization formats like Q8 over floating-point ones like FP8 for better overall fidelity.