Luigi/lfm2.5-350m-cursor-en

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.35BQuant:BF16Context Size:32kPublished:Aug 12, 2026License:otherArchitecture:Transformer Featherless Exclusive Cold

Luigi/lfm2.5-350m-cursor-en is a 350 million parameter Liquid LFM2.5 model, developed by Luigi, specifically fine-tuned as an English-language agent for meeting summarization. It employs a unique CURSOR protocol to stream transcript chunks and emit edit operations for structured, timestamp-anchored meeting notes, rather than traditional map-reduce summarization. This model excels at generating evolving meeting summaries, including decisions, actions, and open questions, and is optimized for on-device deployment with a 4096-token context window.

Loading preview...

Overview

Luigi/lfm2.5-350m-cursor-en is a specialized 350 million parameter Liquid LFM2.5 model, designed for agentic meeting summarization in English. Unlike conventional map-reduce summarization, this model implements a CURSOR protocol where it processes meeting transcripts chunk-by-chunk, generating edit operations to curate a single, evolving set of structured, timestamp-anchored meeting notes. This approach allows the model to track how decisions and topics evolve throughout a meeting.

Key Capabilities

  • Agentic Summarization: Processes transcripts incrementally, revising notes based on new information rather than summarizing independent windows.
  • Structured Output: Generates notes in a defined format (NOTES v2) including TITLE, SUMMARY, DECISIONS, ACTIONS, OPEN items, and TOPICS, each with [m:ss] timestamps.
  • On-Device Optimization: With 350M parameters and a Q4_K_M GGUF size of approximately 215 MB, it is designed to fit within on-device memory constraints.
  • Robust Training: Fine-tuned using teacher traces from Gemma-4-31B, screen-structured synthetic meetings, and real transcripts (QMSum, MeetingBank) to learn the CURSOR protocol and minimize fabrication.

Good For

  • Generating Meeting Notes: Ideal for transforming raw meeting transcripts into organized, timestamped notes that capture decisions, actions, and open questions.
  • Real-time Note Curation: Its chunk-by-chunk processing makes it suitable for applications requiring dynamic, evolving summaries.
  • Edge Device Deployment: Its compact size and efficient architecture (linear attention, 4k context) enable deployment on devices with limited resources.

Limitations

  • English Only: This specific model is trained for English; a separate model exists for Chinese-Traditional.
  • Context Window: Operates with a 4k context per step, handling approximately 2048-token chunks with a state of up to 600 tokens.
  • Synthesized Clocks: Anchors are internally consistent but wall-clock values are synthesized, meaning real audio faithfulness for anchors is unmeasured.

Finetunes
1 models