Kimi K3 reportedly keeps block attention residuals in a 93-layer design
burny_tech · x · 2026-07-28
A repost highlights a video on attention residuals and Kimi K3’s training/infrastructure details.
- Kimi K3 keeps the block attention residuals design unchanged.
- The model uses 8 blocks with 12 layers each, plus a partial final block, or 9 total blocks if the embedding layer is counted.
- It is reported to have 93 total layers and hidden size 7168.
- The speaker argues the model’s aspect ratio of 77 is conservative and reflects a preference for a deeper, not wider architecture, since attention residuals work better in deeper models.
Related event: Kimi K3 shifts attention from scale to architecture(25 posts)→
More from Infra
- AWS tutorial: fleet-wide edge CV model lifecycle with SageMaker SDK v3, ~$0.07 per run — AWS ML Blog · 2026-10-27
- Qwen3.8-27B hits 40 tok/s on Jetson Thor with INT4 + MTP, accuracy beats FP16 — No-Dragonfly6246 · 2026-10-27
- Celesto launches CI runners nearly 90% cheaper than GitHub Actions, with agent sandboxes — aniketmaurya · 2026-10-27
- Supabase founders on rebuilding the backend for a billion AI-built apps — Latent Space · 2026-10-27
- AW-0 runs at 230ms on a single RTX 5090, 2.4x faster than autoregressive code planning — trevordarrell · 2026-10-27
- Site tracks US city council topics: data centers rank only behind pickleball courts — zakkohane · 2026-10-27