Hugging Face releases SmolLM3 mid-training checkpoint amid 200x efficiency debate

eliebakouch · x · 2026-08-20

SmolLM3 author eliebakouch announced the release of the pre-SFT mid-training checkpoint (the it-mid-training branch of SmolLM3-3B-checkpoints, Apache-2.0, 6.17GB), recommending future work start from it — expecting it to be beaten easily given it's a year old. He calls the "200x" claim somewhat overstated but nice work, and agrees mid-training tokens shouldn't be weighted the same as the final SFT dataset.

Original post →

More from Models

Models channel →