Atomic-Germ / LFM2-2.6B-Transcript-NPU2

Try LFM • Documentation • LEAP

LFM2-2.6B-Transcript

Based on LFM2-2.6B , LFM2-2.6B-Transcript is designed for private, on-device meeting summarization . We partnered with AMD to deliver cloud-level summary quality while running entirely locally, ensuring that your meeting data never leaves your device.

Highlights :

• Cloud-level summary quality , approaching much larger models
• Under 3GB of RAM usage for long meetings
• Fast summaries in seconds, not minutes
• Runs fully locally across CPU, GPU, and NPU

Find more information about LFM2-2.6B-Transcript in AMD's blog post and Liquid's blog post .

📄 Model details

Capabilities : The model is trained for long-form transcript summarization (30-60 minute meetings), producing clear, structured outputs including key points, decisions, and action items with consistent tone and formatting.

Use cases :

• Internal team meetings
• Sales calls and customer conversations
• Board meetings and executive briefings
• Regulated or sensitive environments where data can't leave the device
• Offline or low-connectivity workflows

Generation parameters : We strongly recommend using a lower temperature with a temperature=0.3 .

Supported language : English

⚠️ The model is intended for single-turn conversations with a specific format, described in the following.

Input format : We recommend using the following system prompt:

You are an expert meeting analyst. Analyze the transcript carefully and provide clear, accurate information based on the content.

We use a specific formatting for the input meeting transcripts to summarize as follows:

<user_prompt>

Title (example: Claims Processing training module)
Date (example: July 2, 2021)
Time (example: 1:00 PM)
Duration (example: 45 minutes)
Participants (example: Julie Franco (Training Facilitator), Amanda Newman (Subject Matter Expert))
----------
**Speaker 1**: Message 1 (example: **Julie Franco**: Good morning, everyone. Thanks for joining me today.)
**Speaker 2**: Message 2 (example: **Amanda Newman**: Good morning, Julie. Happy to be here.)
etc.

You can replace <user_prompt> with the following, depending on the desired summary type:

This is freeform, and you can add several prompts or combine them into a single one, like in the following examples:

🚀 Quick Start

The easiest way to try LFM2-2.6B-Transcript is through our command-line tool in the Liquid AI Cookbook .

1. Install uv (if you don't have it already):

uv --version
# uv 0.9.18

2. Run with the sample transcript :

uv run https://raw.githubusercontent.com/Liquid4All/cookbook/refs/heads/main/examples/meeting-summarization/summarize.py

No API keys. No cloud services. No setup. Just pure local inference with real-time token streaming.

3. Use your own transcript :

uv run https://raw.githubusercontent.com/Liquid4All/cookbook/refs/heads/main/examples/meeting-summarization/summarize.py \
  --transcript-file path/to/your/transcript.txt

The tool uses llama.cpp for optimized inference and automatically handles model downloading and compilation for your platform.

🏃 Inference

LFM2 is supported by many inference frameworks. See the Inference documentation for the full list.

Name	Description	Docs	Notebook
Transformers	Simple inference with direct access to model internals.	Link
vLLM	High-throughput production deployments with GPU.	Link
llama.cpp	Cross-platform inference with CPU offloading.	Link
MLX	Apple's machine learning framework optimized for Apple Silicon.	Link	—
LM Studio	Desktop application for running LLMs locally.	Link	—

📈 Performance

Quality

LFM2-2.6B-Transcript was benchmarked using the GAIA Eval-Judge framework on synthetic meeting transcripts across 8 meeting types.

Accuracy ratings from GAIA LLM Judge . Evaluated on 24 synthetic 1K transcripts and 32 synthetic 10K transcripts. Claude Sonnet 4 used for content generation and judging.

Inference Speed

Generated using llama-bench.exe b7250 on an HP Z2 Mini G1a Next Gen AI Desktop Workstation on respective AMD Ryzen device. We compute peak memory used during CPU inference by measuring peak memory usage of the llama-bench.exe process executing the command: llama-bench -m <MODEL> -p 10000 -n 1000 -t 8 -r 3 -ngl 0 The llama-bench executable outputs the average inference times for preprocessing and token generation. The reported inference times are for the iGPU, enabled using the -ngl 99 flag.

Memory Usage

Generated using llama-bench.exe b7250 on an HP Z2 Mini G1a Next Gen AI Desktop Workstation with an AMD Ryzen AI Max+ PRO 395 processor. We compute peak memory used during CPU inference by measuring peak memory usage of the llama-bench.exe process executing the command: llama-bench -m <MODEL> -p 10000 -n 1000 -t 8 -r 3 -ngl 0 The llama-bench executable outputs the average inference times for preprocessing and token generation. The reported inference times are for the iGPU, enabled using the -ngl 99 flag

📬 Contact

If you are interested in custom solutions with edge deployment, please contact our sales team .