Our compressed outputs are smaller than WinRaR m5 (highest tier) while being 4-6 faster and using up to 15x less ram during compression, the EXE will be released at 5am UTC September 19, the compressor is currently available online at https://noethix.com/compression we benchmarked it by CPU seconds it takes to compress, both ran on single thread We used GPT-6 Astra to heavily optimize it I would guess about 1B+ tokens were used