Meta: a fine-tuned 2B model hits frontier-level classification at ~1% of the cost, 35ms latency

_philschmid · x · 2026-11-05

Meta's Ahmad Al-Dahle team shares guidance on when to buy versus train your own "decision models": they fine-tuned a 2B model (built on Google's Gemma) for a production classifier, achieving frontier-level quality at 35ms latency and roughly 1% of the cost. They also lay out a framework for when self-training is and isn't worth it; HF's Philipp Schmid praised the work.

Original post →

Meta: a fine-tuned 2B model hits frontier-level classification at ~1% of the cost, 35ms latency

More from Models