dottxt tool-call index shows LFM3-1.7B broken tool calls jump from 7.19% to 15.96% across providers
remilouf · x · 2026-11-18
.TXT (dottxt) published its Tool-call index at providers.dottxt.ai, tracking broken tool-call rates across models and inference providers. Latest finding: Liquid AI's LFM3-1.7B breaks 7.19% of tool calls on Liquid's official fp8 endpoint but 15.96% on Liquid fp8's free tier — the same model, twice the failures. The index covers dozens of models (DeepSeek, Kimi, GLM, Qwen, MiniMax, Mistral, Nemotron, etc.) across providers. Key takeaway: tool-calling reliability depends on the provider and deployment, not just the model — agent builders should factor the serving layer into their selection.
More from coding & agent
- Pocket Folio adds API key + CLI so AI agents can auto-fill your card binders — cnakazawa · 2026-11-18
- Merge Adds Triggers to Agent Handler: Agents Fire Automatically on Emails and PRs — shensi · 2026-11-18
- Frontier models race jets through the Top Gun canyon: Astra wins, Opus refuses to fly — abhi3188 · 2026-11-18
- Running a literature research project inside Grok bot, from question to grant aim — MaxUnfried · 2026-11-18
- Incredibuild spins up Islo, an enterprise 'software factory' for AI agents, with $15M from Insight Partners — Shruti_0810 · 2026-11-18
- MiMo v2.6 Distill flops 0/10 as a coding agent on a 16GB Mac, losing to plain Qwen 9B — GapNew4766 · 2026-11-18