Liquid AI releases on-device meeting summarization model
Liquid AI has released LFM2-2.6B-Transcript, a 2.6-billion-parameter model built with AMD that summarizes hour-long meeting transcripts entirely on-device in under 3GB of RAM, so sensitive meeting data never has to leave the machine.
Liquid AI announced LFM2-2.6B-Transcript, a small model for summarizing long meeting transcripts locally. The company showed it with AMD at CES, running on the AMD Ryzen AI platform across CPU, GPU and NPU. It is available for download on Liquid's LEAP platform and on Hugging Face.
The model is trained on transcripts of 30–60 minute meetings. Users pick an output type with a prompt: executive summary, detailed summary, action items, key decisions, participants or topics discussed. Prompts can also be combined. Transcripts need a specific header (title, date, time, duration, participants) followed by speaker-attributed lines.
Liquid says its configuration processes a 10K-token transcript, roughly a 60-minute meeting, in 2.7GB of RAM. That fits within the ~4GB typically available for AI workloads on a 16GB machine. The company reports turning a 10K-token transcript into a 1K-token summary in 16 seconds.
On accuracy, Liquid says the model beats gpt-oss-20b on short (1K-token) transcripts and comes close to Qwen3-30B and Claude Sonnet. On long (10K-token) transcripts it still beats gpt-oss-20b but falls slightly behind Qwen3-30B and cloud models. These are Liquid's own benchmarks. The post does not name the evaluation method or the hardware used for the speed figure.
According to Liquid, the AMD project went from zero to deployment in under two weeks. A command-line example in the Liquid AI Cookbook runs the model through llama.cpp with no API keys or cloud services.