Estimating K3 Post-Training Costs: ~$4M for the Hero Run
nrehiew_ · x · 2026-07-30
A user performed napkin math to estimate the post-training compute cost for the K3 model, suggesting the total "hero run" cost around $4 million.
Key estimation details:
- The model utilized 50M training rollouts across 24 experts (8 domains × 3 reasoning efforts).
- Each expert received roughly 2M rollouts, covering between 120K to 160K tasks.
- Based on RL plot jaggedness, the author infers fewer than 500 training steps with a relatively large batch size.
- Assuming GB300 NVL72 racks in a colocated setup, the batch size per training replica is likely capped at 1024.
The author notes that this relatively efficient compute usage implies a competent team could potentially push K3's post-training significantly further.
More from Infra
- Testing poolsideai Laguna S 2.1 Inference Acceleration on PGX — gajesh · 2026-07-30
- Qualcomm Seen as the 'Problem Child' of the Current AI Chip Rally — firstadopter · 2026-07-30
- Samsung's Semiconductor Division Operating Profit Soars 24,900% in Q2 — Polymarket · 2026-07-30
- AMD's $1.1M Hackathon: Team Boosts MI355X Performance by 2x via Kernel Optimization — marksaroufim · 2026-07-30
- Samsung Earnings Call Reveals 2nm Projects from Major CSP and HPC Customers — zephyr_z9 · 2026-07-30
- Samsung Q2 Call: Agentic AI Triggers Memory Shortage and Compute Spillover — zephyr_z9 · 2026-07-30