LiquidAI/LFM2.5-350M-Base

TEXT GENERATIONConcurrency Cost:1Model Size:0.35BQuant:BF16Ctx Length:32kPublished:Mar 31, 2026License:lfm1.0Architecture:Transformer0.0K Cold

LFM2.5-350M-Base by LiquidAI is a 350 million parameter pre-trained base model from the LFM2.5 family, designed for on-device deployment. It features a hybrid architecture with 16 layers and a 32,768 token context length, trained on 28 trillion tokens. This model is specifically recommended for heavy fine-tuning in language-specific or domain-specific applications, or for experimenting with novel post-training methods.

Loading preview...

LFM2.5-350M-Base: A Hybrid Model for On-Device Deployment

LiquidAI's LFM2.5-350M-Base is a 350 million parameter pre-trained model, part of the LFM2.5 family, engineered for efficient on-device deployment. Building upon the LFM2 architecture, it incorporates extended pre-training and reinforcement learning, distinguishing it as a hybrid model.

Key Characteristics

  • Architecture: Features 16 layers, combining 10 double-gated LIV convolution blocks and 6 GQA blocks.
  • Training: Pre-trained on an extensive 28 trillion tokens, ensuring a broad knowledge base.
  • Context Length: Supports a substantial context window of 32,768 tokens.
  • Multilingual Support: Capable of processing English, Arabic, Chinese, French, German, Japanese, Korean, Portuguese, and Spanish.
  • Knowledge Cutoff: Information up to mid-2024.

Recommended Use Cases

This base model is primarily intended for scenarios requiring significant customization and adaptation:

  • Heavy Fine-Tuning: Ideal for developing highly specialized models.
  • Domain-Specific Assistants: Creating assistants tailored for particular fields, such as medical or legal.
  • Language-Specific Applications: Adapting the model for specific languages, like Japanese.
  • Proprietary Data Training: Leveraging private datasets for unique applications.
  • Experimental Post-Training: Exploring new approaches in model refinement and optimization.