48M param model achieves tool calling via grammar-constrained design
Turbulent-Sky5396 · reddit · 2026-08-25
Researchers achieved tool calling in a 48M parameter model by designing the tokenizer, loss, and decoder around five specific decisions.
Mechanism & Results:
- The model only handles tool calls; JSON syntax (braces, quotes, keys) is enforced by the grammar, not generated by the model.
- Reliability: Guarantees parseable JSON and prevents undeclared tool calls, eliminating worst-case failures.
- Accuracy: 86.3% on mobile actions and 52.5% on droidcall (in-domain). Unseen catalogs drop to 28 due to function selection errors.
Details:
- Includes a $56 synthetic data script for adapting to custom catalogs.
- Weaknesses: Optional arguments, parallel instances, Java/JS dialects, and 768-token context limit.
- The repo also documents 11 negative results, valuable for similar research.
Related event: 48M-Parameter Model Handles Agent Tool Calls(2 posts)→
More from coding & agent
- Decision models match LLMs on routing and safety — and are 7-8x faster — randal_olson · 2026-10-16
- Temporal Serverless Workers on Cloud Run: queue-driven autoscaling, scale to zero — rseroter · 2026-10-16
- Meta's Superintelligence Lab Runs 200+ Direct Reports—A Lesson for Flat Multi-Agent Systems — peterjliu · 2026-10-16
- LangChain CEO: the most interesting agent harness work lives at the harness-infra boundary — hwchase17 · 2026-10-16
- GitSwarm paper: agents share branchable Git memory, solving all 30 IMOProofBench problems in one run — anirudhg9119 · 2026-10-16
- Hands-on with Gemini 4 Argon: faster than Opus 5.5 but hallucinates more and costs more — prasenx · 2026-10-16