Kimi K3 Full Fine-Tuning Launches on AC2 with 40% Fewer GPUs

Applied Compute launched full fine-tuning for the near-3T-parameter Kimi K3 on its AC2 platform, cutting GPUs per training replica by 40% via memory optimizations including streaming Adam updates and 90GB saved per GPU. Tests show MXFP4 inference error is minimal and matches bf16 training forward results.

2026-08-29 ~ 2026-08-30 · 3 related posts

• Kimi K3 Full Fine-Tuning Live: GPU Requirements Drop by 40% — ypatil125 · 2026-08-29
• Kimi K3 Full Fine-Tuning Launches with Minimal MXFP4 Loss — burny_tech · 2026-08-30
• Kimi K3 Training Optimizations Save ~90GB HBM Per GPU — AccBalanced · 2026-08-30