Liquid AI Publishes LFM3 Collection of Base and Post-Trained Models
Liquid AI has published a Hugging Face blog post titled "Build specialized multimodal models with LFM3", accompanied by a collection of base and post-trained LFM3 models. The listing points to several model sizes and task types, including text generation, image-text-to-text and any-to-any.
- Liquid AI published "Build specialized multimodal models with LFM3" on Hugging Face.
- A linked collection holds base and post-trained LFM3 models, with 7 items.
- Listed models include 2B and 8B text generation, 2B any-to-any, and 2B and 3B image-text-to-text entries.
- The source excerpt gives no benchmarks, licensing or availability details; the full post is at www.liquid.ai/blog/lfm3.
What the LFM3 collection contains
The Hugging Face post is titled "Build specialized multimodal models with LFM3" and appears under the LiquidAI organization. It links to a collection described as a "Collection of base and post-trained LFM3 models". The collection holds 7 items and, at the time of the page snapshot, had been updated 28 minutes earlier. It points readers to www.liquid.ai/blog/lfm3 for the company's own write-up.
The page says 8 models are mentioned in the article. The excerpt available shows only listing metadata, not the full text of the post, so detailed descriptions of each model are not included here.
Sizes and task categories on the listing
The model cards shown on the page cover several categories and parameter counts:
- Text Generation: one 8B model and two 2B models.
- Any-to-Any: two 2B models.
- Image-Text-to-Text: two 2B models and one 3B model.
Several entries were marked as updated within the previous hours, including one listed as updated 5 minutes earlier. This suggests the collection was being filled in as the post went live. The 8B text model and the 3B image-text model carry older update stamps (Aug 24 and 20 days ago respectively), so the listing may also include earlier models that the post references. The page does not say which entries are part of the LFM3 family.
The mix of 2B variants across text, image-text and any-to-any categories matches the post's title, which refers to specialized multimodal models. The source does not give benchmark results, licensing terms, context lengths or hardware requirements, so none are reported here.
Related Liquid AI posts and what it means
The page also lists two earlier posts from the same author. One is "Run fast, private agents on-device with LFM2.5-24B-A2B", dated October 13, 2026. The other is about multimodal open decision models for the edge, dated October 7, 2026, though the title is partly garbled in the source. Both suggest a continuing series of releases from Liquid AI aimed at on-device and edge use, although the source does not spell out that strategy.
For developers, the practical point is that base and post-trained LFM3 checkpoints are grouped in one public collection on Hugging Face. Base models are typically the starting point for custom training, while post-trained versions are meant for direct use, but the source does not elaborate on either. Anyone planning to build on LFM3 will need the full announcement on Liquid AI's site for details on capabilities, performance and terms of use.
Hugging Face: Build specialized multimodal models with LFM3
