Liquid AI runs local tool-calling agent on laptop

Liquid AI reported results from LocalCowork, an open-source desktop agent built on its LFM2-24B-A2B model. The agent runs entirely on-device, with no cloud calls and no API keys. Its source code is in the company's Cookbook.

The company tested the model against 67 tools spread across 13 MCP servers, covering security scanning, audit logging, document processing, file operations, and system and clipboard access. Testing ran on an Apple M4 Max with 36 GB of unified memory, served through llama-server with a Q4_K_M GGUF quantization and flash attention turned on. The workload was 100 single-step tool-selection prompts and 50 multi-step chains of 3–6 steps.

The model averaged about 385 ms per tool-selection response and chose the right tool 80% of the time on single-step prompts. Accuracy was highest in domains with clearly defined tools: 100% on calendar (7/7) and audit (3/3), 90% on security/privacy, and 80% on file operations.

Multi-step tasks were much harder. The model finished only 26% of chains end-to-end. Most wrong picks were "sibling confusion," meaning it chose a related tool on the same server, such as listing files when it should have deleted them. Less often, it answered in plain text instead of calling a tool. Liquid AI says the model works best as a fast dispatcher with a human confirming each step, not as an autonomous agent, and expects targeted post-training to reduce these errors.

The model has been trained on 17T tokens so far, and pre-training is still running. The company's post is cut off before it says what comes after pre-training ends.