Answer.AI releases 33M parameter ColBERT model with outsize performance
lateinteraction · x · 2026-08-19
Answer.AI released a proof-of-concept model, answerai-colbert-small-v1, with just 33 million parameters (MiniLM-sized). Thanks to the new JaColBERTv2.5 training recipe, it outperforms all similarly-sized models on common benchmarks and even beats larger popular models like e5-large-v2 and bge-base-en-v1.5. It is compatible with all recent ColBERT implementations.
More from Models
- Dev shows Qwen3.8 its own README and the model guesses why it performs so well — MikePFrank · 2026-08-19
- Claude Opus 4.6 outputs zero bytes in 900/900 jailbreak trials — rayanpal_ · 2026-08-19
- Qwen 3.8 27B fails complex coding task, lags behind DeepSeek and GLM — myreala · 2026-08-19
- Chart reveals Qwen 3.8 as a massive performance outlier for its size — MikePFrank · 2026-08-19
- Users Complain Claude Code Refuses Geofiltering Tasks — doooyle · 2026-08-19
- Leaked Claude System Prompts Reveal Instruction Evolution from Haiku to Fable 5 — wschroll · 2026-08-19