ErtasAI/LFM2.5-350M

TEXT GENERATIONConcurrent Unit Cost:1Model Size:0.35BQuant:BF16Context Size:32kPublished:Jul 16, 2026License:lfm1.0Architecture:Transformer Featherless Exclusive Cold

LFM2.5-350M is a 350 million parameter hybrid model developed by Liquid AI, designed for efficient on-device deployment. It features extended pre-training on 28T tokens and large-scale multi-stage reinforcement learning, enabling high-quality AI in a compact size. This model excels at fast edge inference, supporting various platforms like llama.cpp, MLX, and vLLM, and is optimized for data extraction, structured outputs, and tool use.

Loading preview...

LFM2.5-350M: Optimized for On-Device AI

LFM2.5-350M, developed by Liquid AI, is a 350 million parameter hybrid model built upon the LFM2 architecture. It is specifically engineered for on-device deployment, offering high-quality AI performance in a compact footprint. The model benefits from significantly extended pre-training, now on 28 trillion tokens, combined with large-scale multi-stage reinforcement learning.

Key Capabilities & Features

  • Best-in-Class Performance for Size: Achieves performance comparable to much larger models, making advanced AI accessible for edge devices.
  • Fast Edge Inference: Delivers rapid decode speeds (e.g., 313 tok/s on AMD CPU, 188 tok/s on Snapdragon Gen4) and operates under 1GB of memory.
  • Broad Platform Support: Offers day-one support for llama.cpp, MLX, and vLLM, with optimized formats for various hardware including Apple Silicon and Intel platforms.
  • Tool Use: Supports function calling with a flexible mechanism for defining tools via JSON in the system prompt and interpreting tool outputs.
  • Multilingual Support: Capable in English, Arabic, Chinese, French, German, Japanese, Korean, Portuguese, and Spanish.

Use Cases & Recommendations

LFM2.5-350M is particularly well-suited for:

  • Data Extraction
  • Structured Outputs
  • Tool Use / Function Calling

It is not recommended for knowledge-intensive tasks or programming. The model's design prioritizes efficiency and performance for constrained environments, making it an excellent choice for integrating AI directly into applications and devices.