Stability AI AI Updates: October 31, 2026

1. Stability AI Fixes Audible Artifacts in the Quantized CPU Codec for Stable Audio 3 Optimized

Stability AI pushed two rounds of fixes to its stable-audio-3-optimized repository, the experimental collection of compiled Stable Audio 3 runtime artifacts (TensorRT plans for sm_90 GPUs and TFLite/LiteRT codecs for CPU). On October 29 the w8a8 SAME-S decoder was rebuilt with per-token output-mapping quantization to remove a 172 Hz buzz, and on October 30 the w8a8 SAME-S encoder gained a rotation (labeled RHT in the commit) on the block-0 GLU projection that lowers its noise floor by about 11 dB at the same size and speed. The previous artifacts are archived under legacy/. The LiteRT codecs require ai_edge_litert 2.2.0 or newer and are what make CPU-only generation possible; the model card cites a few seconds per clip on a MacBook Pro M4 and under 2 seconds on an H200 with the TensorRT engines, which target sm_90 only and must be rebuilt locally for other architectures (about 5 to 8 minutes). The artifacts are under the Stability AI Community License, with the T5Gemma text encoder redistributed under the Gemma Terms of Use. Source Source