Oriel — AI & technology worth noticing, every day.

Build specialized multimodal models with LFM3

Web page · https://huggingface.co/blog/LiquidAI/lfm3

First seen by Oriel on .

AI summary

Liquid AI introduces LFM3, a family of specialized 2B-class text, vision, audio, omni, and decision models sharing an edge-oriented backbone that can be fine-tuned on one GPU.

  • LFM3-1.7B-Base uses 13 gated convolutions, five sliding-window attention layers, and six global attention layers, with a 256K-token context window.
  • The backbone was pre-trained on 22T tokens, including a final 3.5T-token stage emphasizing curated agentic traces and tool calls.
  • The family includes LFM3-1.7B, LFM3-VL-2B, LFM3-Audio-2B, LFM3-Omni-2B, d1-2B, and d1-omni-2B, all supported through transformers.
  • LFM3-VL-2B averages 71.4 across 62 vision benchmarks, while its 164M-parameter vision encoder requires 2.2x fewer FLOPs than Qwen3.5-2B's encoder.
  • The text model occupies about 1 GB at 4-bit precision, while LiquidCodec streams speech and decodes around 87x real time on one Apple M4 Max CPU thread.

#multimodal models #edge ai #transformers #model fine-tuning

Generated by gpt-5.6-sol. Check the original for details.

Aliases

SourceExternal IDURL
Hugging Face Bloghttps://huggingface.co/blog/LiquidAI/lfm3https://huggingface.co/blog/LiquidAI/lfm3 Read discussion