Liquid AI's New Model Matches DeepSeek in Agentic Tool Calls at 3.7x Speed
JosephJacks_ · x · 2026-08-07
Liquid AI's newly released 2.6B parameter model, LFM2.5-2.6B, achieves a breakthrough in agentic tool-calling capabilities. Tested against DeepSeek-V4-Flash on a machine with 4x RTX 5090s, the results highlight significant efficiency gains.
- Tasks: Included checking weather/time in six cities, currency conversion, and hotel booking. These tasks require the model to successfully fire all necessary tool calls to complete.
- Performance: Both models successfully completed all 35 tool calls, with 12 fired in a single turn. However, LFM2.5-2.6B operated at 366 tokens/s, finishing in 19 seconds, compared to DeepSeek-V4-Flash at 77 tokens/s and 70 seconds.
This demonstrates the efficiency advantages of Liquid's agent-optimized model for edge deployment (like smartphones) and high-frequency tool-calling scenarios.
More from coding & agent
- Cua Driver Launches Extension-Free Browser Use Interface for AI Agents — rudrank · 2026-08-07
- LangChain Releases Guide on Open Source Agent Frameworks — LangChain · 2026-08-07
- Naïve Raises $28.5M Series A to Build Business Infrastructure for AI Agents — Scobleizer · 2026-08-07
- AI Coding Practice: Solid Planning Does 80% of the Job — EXM7777 · 2026-08-07
- HuggingFace CEO: Public Agent Collaboration Makes AI Safer and More Efficient — ben_burtenshaw · 2026-08-07
- Don't Let Coding Agents Grade Their Own Homework: Introducing QA Agents — sergeykarayev · 2026-08-07