lazos/lfm2.5-350m-frontend-agent

TEXT GENERATIONConcurrent Unit Cost:1Model Size:0.35BQuant:BF16Context Size:32kPublished:Jul 14, 2026License:lfm-open-license-1.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The lazos/lfm2.5-350m-frontend-agent is a 350 million parameter English web/front-end agent, fine-tuned from LiquidAI/LFM2.5-350M, designed to run entirely in the browser. It operates as a stateless, context-bounded agent, grounding each turn in a compact context block provided by the host application, enabling precise interaction with on-screen elements like views, carts, and knowledge. This model excels at interpreting and acting upon dynamic web content, making it suitable for interactive frontend applications where a small, efficient agent is required.

Loading preview...

Model Overview

The lazos/lfm2.5-350m-frontend-agent is a specialized 350 million parameter language model, fine-tuned from LiquidAI/LFM2.5-350M, engineered for in-browser web/front-end agent applications. It is designed to be small enough to run efficiently on the edge using technologies like wllama, without requiring server-side processing.

Key Capabilities & Features

  • Context-Bounded Operation: The agent is stateless and grounds every interaction in a compact context block (VIEW, CART, KNOWLEDGE) provided by the host application, ensuring actions are based solely on current, explicit information.
  • Precise Interaction: It accurately interprets and acts upon displayed items, including adding/removing by ID, resolving references (e.g., "the second one"), and answering price queries based on the current view.
  • Enhanced Resolution (v1.3.0): Improvements in version 1.3.0 allow for searching items by name first for those not currently on screen, and better handling of filtering conditions even when items are already in view.
  • Off-Scope Steering: The model can gracefully decline genuinely off-topic requests and gently guide users towards its capabilities, while also engaging in brief, in-character small talk.
  • Quantity Disambiguation: It avoids misinterpreting numbers embedded in titles as quantities.
  • Open-License Training Data: Trained on synthetic data derived from Apache-2.0 / MIT licensed teacher models (Qwen2.5-7B + Qwen3-30B-A3B), ensuring no proprietary model output was used.

Ideal Use Cases

This model is particularly well-suited for:

  • Browser-based AI Assistants: Implementing interactive agents directly within web applications for tasks like e-commerce navigation, form filling, or content interaction.
  • Edge Computing: Deploying AI capabilities directly on user devices for low-latency, privacy-preserving interactions.
  • Frontend Automation: Creating agents that can understand and respond to user queries about on-screen elements and perform actions like adding items to a cart or filtering lists.

For the smallest footprint, a 230M parameter variant (lazos/lfm2.5-230m-frontend-agent) is also available, offering a trade-off between size and capability.