AICoven Qwen3 1.7B MCP
A fine-tuned Qwen3 1.7B model optimised for MCP (Model Context Protocol) tool calling in AICoven.
Training Details
- Base model:
mlx-community/Qwen3-1.7B-4bit - Method: LoRA fine-tuning (16 layers, rank 16)
- Training examples: 284 (190 tool calls, 59 no-tool, 35 multi-turn)
- Iterations: 300, learning rate 1e-4
- Final validation loss: 0.011
- Accuracy: 87.9% on 58 test examples (tool routing)
Capabilities
- Routes user requests to the correct tool from 75+ MCP tools
- Handles file operations, shell execution, GitHub, calendar, Slack, Notion, and more
- Answers general knowledge questions directly without tool calls
- Processes multi-turn conversations with tool results
Usage
This model is designed for use with the AICoven macOS/iOS app using the MLX framework.
from mlx_lm import load, generate
model, tokenizer = load("aicoven/Qwen3-1.7B-MCP")
- Downloads last month
- 6
Model size
0.3B params
Tensor type
BF16
·
U32 ·
Hardware compatibility
Log In to add your hardware
4-bit
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support