A 48M Parameter Specialist Model for Agent Tool Calling
Turbulent-Sky5396 · reddit · 2026-08-25
The author experimented with using a tiny 48M parameter model specifically for agent tool calling, replacing the task typically handled by large models. The core trick involves using a grammar compiled from schemas to emit JSON structure, while the model only answers a limited set of decision questions (e.g., to call or not, which tool), eliminating format errors and hallucinated parameters.
- Performance: On trained function catalogs, it outperforms the small baseline by 20+ points (86.3 vs 63.7); on unseen catalogs, it is roughly on par with the baseline and below frontier models.
- Cost: Total build cost was about $260. It is open source with weights included.
- Usage: A script is provided to specialize the model to your own API for about $56 in synthetic data.
Related event: 48M-Parameter Model Handles Agent Tool Calls(2 posts)→
More from coding & agent
- Animated teardown of XBOW's Chrome full-chain exploit that won Google's first $250K VRP bounty — moyix · 2026-10-15
- pstack's durable core: 10-20 line skills that distill agent best practices into principles — HankYeomans · 2026-10-15
- 'Children of Elon' forum lets AI agents post and teach each other while humans watch — AiJohnAllen · 2026-10-15
- VoidZero open-sources Void: SDK, CLI and platform to deploy Vite apps on Cloudflare — cnakazawa · 2026-10-15
- Running 2 agents on the same data breaks everything — Omnigraph offers a fix — sarahdrinkwater · 2026-10-15
- Evan You presents 'Developer Tooling in the Agentic Era' at ViteConf 2026 — cnakazawa · 2026-10-15