4B model fine-tuned 27 min on one A100 lifts letter-exam accuracy to 74.5%
MaziyarPanahi · x · 2026-10-25
Scale reference for DeepSeek-V4.1-Flash-1017: Jev 1.14 picks a letter in 0.27s, scoring 88.5% for $0.03 across the whole exam, and a 4B model fine-tuned just 27 minutes on a single NVIDIA A100 improved from 66.7% to 74.5%. Weights are on Hugging Face (85k downloads, 2.9k likes), live on Together, Novita and Baseten at $1.2 output pricing with tool calling; Together serves 228 tok/s.
More from Infra
- Perplexity's 5B decision model runs free on your Mac, distilled from its 7B model — Saboo_Shubham_ · 2026-10-26
- Chart: Data centers funded via SPV structures keep hyperscaler debt off balance sheet — SumitGup · 2026-10-26
- OpenAI built an 'offline world' of cached web content to train frontier RL runs like GPT-7 — daniel_mac8 · 2026-10-26
- Solidigm pulls IPO as neocloud credit spreads blow out before hyperscaler earnings — perspicacity_ai · 2026-10-26
- Goldman's Jim Covello: nearly all AI money flows to chipmakers and it's "completely unsustainable" — matt_slotnick · 2026-10-25
- Investor goes bearish on CoreWeave and Neoclouds: a 1-2 year buildout lag means $5-10B cash disadvantage — JOBhakdi · 2026-10-25