Fine-tuned open-source models beat frontier models, cut costs 95%
pdamodaran · x · 2026-11-03
From Palantir's earnings call: L3 Harris fine-tuned open-source models on its own data and, in preliminary results, outperformed its frontier model in under 48 hours while cutting costs 95%. Nvidia post-trained its own open models so the smallest one beat its largest by a wide gap and can now retrain on a single GPU in hours. Tampa General is deploying local sovereign AI with fine-tuned Nematron to keep sensitive data on-prem and enforce its own LLM risk and privacy controls.
More from Infra
- Periodic Labs to detail custom RL infra for physical experiments at SGLang Summit 2026 — hsu_byron · 2026-11-03
- NVIDIA Research Blog: Why KV Cache Compression Breaks in Production vLLM Deployments — blaizedsouza · 2026-11-03
- mlx-learn launches: scikit-learn API on Apple silicon with up to 30X training and 300X inference speedups — HankYeomans · 2026-11-03
- Podcast episode: Is running local AI for you? — rez0__ · 2026-11-03
- ByteDance crawler spotted scraping websites at massive scale — gefei55 · 2026-11-03
- 1e29 FLOPS would take 200k GB200s 3.2 years and 6.2 TWh, dev notes — mike64_t · 2026-11-03