SmolDuplex: Full-Duplex Interaction Model (135M)

A sub-200M parameter full-duplex spoken interaction model with 200ms turn-taking, built on SmolLM2-135M + CosyVoice. See ARCHITECTURE.md for the complete PRD.

Key Specs

• Trainable params: ~139M
• Turn-taking latency: <400ms
• Chunk size: 200ms (13 tokens per chunk)
• Capabilities: ASR, TTS, simultaneous listen+speak, backchanneling, interruption handling
• Training cost: ~$44 cloud / ~$2 own GPU
• Training time: ~38 GPU-hours
• Inference: Any 8GB+ GPU, real-time

Based On

• OmniFlatten — proven full-duplex at 500M
• SyncLLM — time-synchronized chunks
• SmolLM2-135M — LLM backbone
• CosyVoice — speech tokenizer/detokenizer

Generated by ML Intern

This model repository was generated by ML Intern, an agent for machine learning research and development on the Hugging Face Hub.

• Try ML Intern: https://smolagents-ml-intern.hf.space
• Source code: https://github.com/huggingface/ml-intern

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "PranavHarshan/SmolDuplex"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

For non-causal architectures, replace AutoModelForCausalLM with the appropriate AutoModel class.