Thinking Machines details its NVFP4 training recipe: per-expert scales, 4/6 activation quantization
snowclipsed · x · 2026-11-14
Screenshots from Thinking Machines' blog break down the NVFP4 recipe behind Inkling 2 Small: the last 16.6% of layers stay in higher precision (MXFP8 instead of BF16); activation quantization uses Four Over Six (4/6), picking between two scale factors to cut quantization error, while weights skip 4/6. Borrowing from Composer 2, it applies per-expert outer scales for weights and per-token outer scales for activations, giving efficiency, batch-invariance, and no causality leakage. The recipe ran unchanged through SFT and RL, costing 2% compute efficiency in pretraining versus MXFP8 in exchange for NVFP4 speedup, while RL saw a substantial end-to-end speedup with no discernible reward difference.
Related event: Thinking Machines Details Native 4-Bit Training Recipe for Inkling 2 Small(4 posts)→
More from Infra
- Grok confirms July 2024 CrowdStrike update grounded major US airlines nationwide — TinfoilTricorn · 2026-11-14
- National Weather Service hit by nationwide internet and communications outage — MicahBerkley · 2026-11-14
- Polymarket odds put 39% on FAA outage being attributed to AI — Polymarket · 2026-11-14
- Redditor predicts Chinese AI chips will flood the market as Nvidia stumbles — Apart_Tumbleweed_303 · 2026-11-14
- AWS shows how to route coding agent requests by cost per solved task with NeMo Switchyard on Bedrock — AWS ML Blog · 2026-11-14
- GAO: 76% of US air traffic control systems unsustainable, still on floppies and Win95 — MartinGTobias · 2026-11-14