Mirendil reports one researcher pretrained Mirendil-Base in 3.5 months for under $10m in compute, with agents writing over 99% of code; the company says post-training has begun, not autonomous self-improvement.

Mirendil’s original held-out-code loss chart includes unfinished Large/XL runs and base-versus-post-trained comparison caveats; company research results, not an independent benchmark.

Published 21 Oct 2026 · 16:40 UTC

Mirendil ↗ · Source dated 21 Oct 2026

Browse all Live updates · Read Features