CAAL Qwen3.5 2B — Fine-Tuned for Tool Calling

A fine-tuned Qwen3.5 2B model optimized for tool calling in voice assistant workflows. Built for CAAL (CoreWorxLab Ambient Assistant for Linux).

Performance

82/85 tests passed (96%) on the CAAL 85-test evaluation suite:

Category	Score
Single tool calls	27/27
Conversational (no tool)	10/10
Multi-turn chains	28/29
Argument formatting	17/19

Model Details

Base model: Qwen3.5 2B
Training method: SFT with BF16 LoRA (last-turn-only — previous turns as context, only final response trained)
LoRA config: r=32, alpha=32
Quantization: Q4_K_M (GGUF)
File size: ~1.2 GB
VRAM usage: ~2.6 GB at 16384 context

Usage with Ollama

# Download the GGUF and create a Modelfile:
# Modelfile contents:
# FROM caal-qwen3.5-2b-q4.gguf
# RENDERER qwen3.5
# PARSER qwen3.5
# PARAMETER temperature 0.1
# PARAMETER num_ctx 16384

ollama create caal-qwen35-2b -f Modelfile

Designed For

Edge deployment on consumer GPUs (fits on 5GB+ VRAM alongside TTS)
Local voice assistants with tool calling
Smart home control, email, calendar, and service management
Multi-step tool chains (e.g., search → lookup contact → send email)

License

See LICENSE for the CAAL Model License v1.0. This model is free for personal, non-commercial use with attribution to CoreWorxLab. Commercial use requires written permission.

The base model (Qwen3.5) is licensed under Apache 2.0. Users must comply with both licenses.