Luigi/voxsum-gemma3-270m-litertlm
The Luigi/voxsum-gemma3-270m-litertlm is a 270 million parameter Gemma-3 based model, developed by Luigi, specifically fine-tuned for generating structured meeting notes from zh-TW and English transcripts. It produces summaries, decisions, action items, open questions, and topics with timestamps, optimized for long context (32,768 tokens) despite its small size. This model is designed for on-device deployment, offering a compact solution for meeting summarization.
Loading preview...
Model Overview
Luigi/voxsum-gemma3-270m-litertlm is a compact 270 million parameter model, based on Gemma-3, designed for generating structured meeting notes. It is specifically fine-tuned for both traditional Chinese (zh-TW) and English transcripts, producing summaries, decisions, action items, open questions, and topics, each with [m:ss] timestamps for easy verification against audio. The model is packaged as a .litertlm file for LiteRT-LM, with the primary evaluated build being 501 MB, and a smaller 159 MB variant available for size/latency experiments (though its quality is unevaluated).
Key Capabilities
- Structured Meeting Notes: Generates comprehensive notes including summaries, decisions, action items, open questions, and topics.
- Timestamped Output: Each bullet point in the notes includes a timestamp, allowing users to quickly reference the original audio.
- Multilingual Support: Fine-tuned for both zh-TW and English meeting transcripts.
- Long Context Handling: Utilizes a 32,768-token KV cache and 2,048-token prefill chunks, leveraging Gemma-3's sliding-window attention and MQA for efficient long-context processing at a small size.
- Quantization-Aware Training (QAT): Trained with int4 quantization, making it optimized for low-precision inference and memory efficiency.
Use Cases and Differentiators
This model is particularly suited for on-device applications where memory and speed are critical constraints. While a larger voxsum-qwen3-0.6b model offers better faithfulness and coverage, voxsum-gemma3-270m provides a significantly smaller footprint (501 MB vs 664 MB) and comparable inversion rates (12.5%). It excels in scenarios requiring efficient, localized meeting summarization, especially for zh-TW transcripts where it shows 0% inversion rates. The model's training methodology, using a same-family teacher (Gemma-3-27B-it) and a single-pass method, significantly reduces contradictions and empty content compared to other approaches. For very long documents, a wide-window map-reduce strategy with the 32k build is recommended over a harness-trained approach, yielding better results with fewer LLM calls.