Overview & Result
Marius Vach describes pairing Jev with LiquidAI's LFM 2-350M-extract in a personal assistant: Jev decides which tool to call, while the smaller extraction model produces the arguments.
How Jev fits in the loop
- The assistant receives the user query and current state.
- Jev selects one tool from the bounded tool set.
- LFM 2-350M-extract turns the query into arguments for that tool.
How to reproduce
- Open the original post and preserve the two-model division it describes.
- Give Jev a bounded tool set and let it choose the next tool.
- Pass the selected tool and user query to LFM 2-350M-extract for argument extraction, then execute in code.
Why this build matters
Choose the action with Jev, extract parameters with a small generative model, and keep execution deterministic.
Limitations
- The personal assistant is described as in progress; no repository or completed benchmark is linked.
- The post does not publish the tool schema, confidence thresholds, or failure handling.