Skip to main content

General Questions

LFM (Liquid Foundation Models) are Liquid AI’s efficient models for local and hosted deployment. LFM3 is the latest generation, with text, vision, audio, omni, and decision models. The library also includes smaller and specialized LFM2.5 and LFM2 models, including embedding and encoder models. Open-weight models range from 230M to 24B parameters.
LFM3-1.7B supports 256K tokens, and LFM3-VL-2B supports 128K. Other limits vary by model: many LFM text, vision, and audio models support 32K, LFM2.5-8B-A1B supports 128K, and LFM2.5-24B-A2B supports 131K. Encoder-based models have shorter limits (8K for the LFM2.5 encoders, 16K for d1-omni-600M). Each model page lists its exact limit; hosted API limits can differ from the open-weight checkpoints.
LFM models are compatible with the frameworks below. Support varies by model (for example, audio and decision models run on fewer frameworks), so check each model page.

Model Selection

  • LFM3 is the latest generation. Its tri-hybrid architecture combines gated short convolutions, sliding-window attention, and global attention. The family includes text, vision, audio, omni, and decision models, with updated chat templates and tool-call formats.
  • LFM2.5 builds on the LFM2 architecture with improved training. Some models extend the context window or vocabulary. It remains useful for smaller footprints, larger MoE models, and specialized tasks.
  • LFM2 is the earlier generation. Some specialized models remain useful; deprecated models and their replacements are listed on the Deprecations page.
Start with LFM3 for general use, then consider an older model when its size, specialization, or runtime support better fits your needs. Use the selected model’s own chat template, tool-call format, and generation settings.
Liquid Nanos are task-specific models fine-tuned for specialized use cases like:
  • Structured extraction from images (LFM2.5-VL-1.6B-Extract, LFM2.5-VL-450M-Extract)
  • Retrieval for RAG pipelines (LFM2.5-Embedding-350M, LFM2.5-ColBERT-350M)
  • Translation (LFM2-350M-ENJP-MT)
  • Meeting summarization (LFM2-2.6B-Transcript)

Deployment

Yes! llama.cpp runs natively on iOS and Android with quantized GGUF models. See iOS & Android for embedding it in an app, and Vision & Audio for multimodal models.
  • GGUF: For llama.cpp, LM Studio, Ollama, Atomic Chat (Q4_0, Q4_K_M, Q5_K_M, Q6_K, Q8_0, BF16, F16; some repositories ship a subset)
  • MLX: For Apple Silicon (4-bit, 5-bit, 6-bit, 8-bit, bf16)
  • ONNX: For cross-platform deployment with ONNX Runtime
  • Q4_0 / 4-bit: Smallest size, fastest inference, some quality loss
  • Q8_0 / 8-bit: Good balance of size and quality
  • F16 / bf16: Full precision, best quality, largest size
For most use cases, Q4_K_M or Q5_K_M provide good quality with significant size reduction.

Fine-tuning

Yes. Most open-weight LFM models support fine-tuning with TRL and Unsloth, MoE models can be trained across GPUs with Halo, and audio models use the liquid-audio trainer. Check the Model Library for trainability information.
  • LoRA/QLoRA: Memory-efficient fine-tuning
  • Full fine-tuning: For maximum customization
  • SFT (Supervised Fine-Tuning): For instruction tuning

Still Have Questions?