Palantir: fine-tuned 4B Nemotron beats GPT-6 Astra and Claude Sonnet 5.5 at 1/200th the cost
eliano · x · 2026-11-03
In its Q3 business update, Palantir showed that a fine-tuned NVIDIA Nemotron 3 Nano (4B params) beat much larger frontier models on a real RFI-classification task for construction firm McCarthy:
- Nemotron 4B fine-tuned: 75% accuracy at $0.10/M tokens
- Claude Sonnet 5.5: 66% at $4/M
- GPT-6 Astra: 62% at $20/M
The untuned 4B base scored only 53%, so fine-tuning added roughly 20 points. Palantir estimates 30,000 drafting hours saved annually for McCarthy. NVIDIA is also training open models on its own supply chain data (millions of parts, thousands of suppliers) inside Palantir's Ontology, with Jensen Huang framing the partnership as turning operational graphs into sovereign intelligence.
More from Infra
- A single B300 now matches the world's fastest 1997 supercomputer in FP64 FLOPs — scaling01 · 2026-11-03
- Anthropic IPO: investors push for $1.5T-$1.6T valuation, below ~$2T discussed — rohanpaul_ai · 2026-11-03
- Cerebras engineer explains wafer-scale cooling: water flows straight into the wafer — bookwormengr · 2026-11-03
- Goertzel counters Altman's idle-RAM take with decentralized compute networks — bengoertzel · 2026-11-03
- Satire: companies borrow a 'quadrillion' to build 420 data centers in 6 months — asymco · 2026-11-03
- SF billboards slam cloud AI for stealing data; blogger predicts decade of on-device iPhone assistants — brandon_xyzw · 2026-11-03