LFM2.5-1.2B-Instruct
LiquidAI/LFM2.5-1.2B-Instruct is positioned as a “real” assistant model for edge devices: small enough to run locally, but trained as an instruction-tuned chat model (and not just a toy demo). Liquid AI emphasizes on-device deployment and broad runtime support (Transformers, vLLM, llama.cpp, and MLX), which makes it easy to benchmark the same checkpoint across different local inference stacks.
What’s useful about this upload is how explicit it is about intended use: the card calls out agentic tasks, data extraction, and RAG-style workflows, and (importantly) warns that it’s not recommended for knowledge-heavy queries or programming. The model also advertises a relatively long context window for its size (32,768 tokens) and a multilingual vocabulary, which is a practical combo for “local document assistant” experiments.
If you want to try it quickly, start with the model’s ChatML-like template (via tokenizer.apply_chat_template() in Transformers) and test a few structured prompts: extraction into JSON, short summarization, and tool-style function calling. Those are exactly the places where small, fast instruction models can feel surprisingly capable.
Quick stats from the listing feed: pipeline: text-generation · 121 likes · 2917 downloads.
Source listing: https://huggingface.co/models?sort=modified