OneNeuralX1Tool - 26M Function Calling Model
26M parameter encoder-decoder for tool calling.
Architecture
- Encoder: 12 layers hybrid attention (75% linear, 25% full)
- Decoder: Recurrent (4 loops) + ACT halting
- MoE: 4 experts + shared expert
- Vocab: 8,192 BPE
Training
- Data: Needle tool-calling dataset
- Accuracy: 85.83%
- Epochs: 20
Tools
search_web, get_weather, send_email, create_event, get_stock_price, set_reminder, play_music, calculate
- Downloads last month
- 26
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support