Hands busy? A 4B AI kitchen assistant that sees, talks and uses tools

Wet hands, a busy stove, and a phone you would rather not touch. We built a short kitchen assistant prototype with DRVEXO-4B Unified.

1. Show the ingredients and state the constraints.
   The example uses tomatoes, eggs and cucumber, with a request for two people, under 20 minutes, light seasoning and no soup. The model calls search_recipes with those constraints and chooses tomato-and-egg stir-fry with cucumber salad from three returned recipe candidates.

2. Start a three-minute timer by voice.
   The model generates start_timer(seconds=180). The local tool creates the timer record and returns the result for a spoken confirmation.

3. Add milk to the shopping list, without buying it.
   The next call writes milk to the local shopping list. No checkout or payment step is involved.

4. Return to the original timer.
   After the shopping-list turn, the assistant calls get_active_timer and reads the same timer record. The captured response is 2 minutes 54 seconds remaining.

The 41-second animation was made in Manim, with an abstract assistant sphere, speech-driven waves and visual tool receipts. The model responses and tool calls come from the recorded run; the recipe selection is within three candidates, and the tools run in a local prototype. DRVEXO uses a 4B language backbone.

Which hands-busy task would you want an assistant to handle?

Hugging Face: https://huggingface.co/DRVEXO-LLM/drvexo-4b-unified
ModelScope: https://modelscope.cn/models/DRVEXOLLM/drvexo-4b-unified