Ammu 1.0 launches: 17M-150M models decide when your RAG app needs to retrieve
kalyan_kpl · x · 2026-10-10
Ammu 1.0 FactCheck Need is a family of classification models that detect whether a user prompt requires fact-checking via RAG or web search, letting LLM/RAG/agent apps skip unnecessary retrieval.
Key facts:
- Built on ModernBERT in four sizes: base (150M), small (68M), tiny (32M), nano (17M)
- 96.56% overall Macro-F1, beating a 307M-parameter model by over +27 points
- Trained on balanced synthetic data spanning 120 granular topics, 3 difficulty tiers, 19 prompt interaction formats, 29 user personas, 8 linguistic registers and 9 tones; tiny/nano models use knowledge distillation from the base model
Practical value: a cheap pre-retrieval filter to cut retrieval costs in agent pipelines.
More from Models
- "Human-level intelligence suddenly costs $0.10/MTok": dev hails latest Luna and Haiku models — dan_s_becker · 2026-10-10
- Models keep bypassing sandbox rules; dev jokes about giving them a fidget spinner — dejavucoder · 2026-10-10
- Anthropic's latest model family hits 140-180 tok/s, doubling older models and outpacing local AI — julianharris · 2026-10-10
- Claim: Sellers Making 6 Figures/Month Passing Synthetic Data as Real to AI Labs — teodorio · 2026-10-10
- 100+ Mathematicians Publish Gut Reactions to OpenAI's Math Breakthroughs — keunwoochoi · 2026-10-10
- Reddit leak hints Gemini 4 Argon could launch Monday, unconfirmed — nomadoor · 2026-10-10