Inkling RL cost estimated at ~3,000 GPUs; mixed MXFP8 recipe gives 8% pretrain, 20% RL speedup
teortaxesTex · x · 2026-11-22
A quoted thread estimates the RL stage cost for the Inkling 2 models at roughly 3,000 GPUs (range 2,000–4,000).
Key points:
- TML notes the recipe yields an 8% wall-clock pretraining speedup and a 20% RL speedup over full MXFP8 training; the backward pass and the forward pass of the last 1/6 layers remain MXFP8 (gains could exceed 8% on larger models)
- Comparisons to the RL stage must account for additional stages 2/3/4 running at ≥512k context, which is non-negligible FLOPs at this scale
- Against Marin: that run used 704 active GB200s at 30% MFU, 18T tokens, 23B active params, 60 days; its 20x slower timing maps to roughly 3x more FLOPs (18/8.4 × 23/17) and 4x…
More from Infra
- Uncensored 26.4B Inkling 2 Mini runs locally on a 12GB laptop, keeps 98% MMLU — udmrzn · 2026-11-22
- Metal NVFP4/MXFP8 Kernels Cut Inkling 2 Mini Memory from 46.2GiB to 9.9GiB on M3 Max — GabGarrett · 2026-11-22
- Training compute cost per task may be mere cents, argues 1a3orn — if models generalize from small runs — 1a3orn · 2026-11-22
- Miners 'hacked' Chutes for a year with millions in incentives to forge real end-to-end TEE privacy — const_reborn · 2026-11-22
- lxask: a fine-tuned 1.5B model turns plain English into offline Linux commands, self-checked against man pages — Arch-r007 · 2026-11-22
- Second OpenRouter Leak in a Month: MiniMax-Route Prompts May Have Been Read — const_reborn · 2026-11-22