4B specialized model hits 93.8% on bank support, beating proprietary models' 74%
MParakhin · x · 2026-11-14
Toloka trained a 4B model for bank support, where every reply must follow strict policy: test conversations solved went from 1% to 93.8%, while proprietary models like Gemini 3.8 managed only 74%.
Mikhail Parakhin comments that the pattern of "develop with the largest model at highest reasoning, then build a specialized targeted model for production" is gradually percolating through the industry — though he's surprised adoption isn't higher, as he constantly sees people doing the reverse.
More from coding & agent
- Metabase MCP Server ships 33 tools letting LLMs query dashboards and run SQL — modelcontextprotocol · 2026-11-14
- Harbor launches rollout servers: a control plane for agent workloads — WalterReade · 2026-11-14
- Semgrep CEO: frontier models now write working exploits — sandbox your agents, enforce with hooks — blaizedsouza · 2026-11-14
- fofrAI adds realistic human motion to custom VFX tool using Carbon agents — fofrAI · 2026-11-14
- Paper argues 'The End of Harness Engineering': loops rewrite agent harnesses, humans just write the exam — blaizedsouza · 2026-11-14
- Salnate adds credential management to its AI teammates, with shared or dedicated logins — inductionheads · 2026-11-14