LiquidAI/LFM2.5-Encoder-230M-GGUF

Try LFM • Documentation • LEAP

LFM2.5-Encoder-230M

LFM2.5-Encoder-230M is a multilingual bidirectional encoder built on the LFM2 architecture — a lightweight encoder for tight latency and memory budgets, punching above its size class. It is a masked language model with full bidirectional attention, designed to be fine-tuned into task-specific models (classification, token classification, retrieval, reranking, and semantic similarity) across 15 languages, and to run efficiently on-device.

• Highly capable for its size. On par with the best similarly sized encoders and well ahead of our own retrieval siblings.
• General-purpose. 8k context, strong across NLI, paraphrase, sentiment, and multilingual tasks.
• Fast and on-device. Matches or beats ModernBERT throughput, with a long-context edge on CPU; runs in the browser on WebGPU.

Find more information about LFM2.5-Encoder-230M in our blog post.

🏃 How to run

Example usage with llama.cpp:

Start llama-server with per-token embeddings

hf download LiquidAI/LFM2.5-Encoder-230M-GGUF LFM2.5-Encoder-230M-F16.gguf --local-dir .
llama-server -m LFM2.5-Encoder-230M-F16.gguf --embeddings --pooling none

Run masked-token prediction — the mask position's logits come from the per-token hidden states and the tied embedding matrix read from the GGUF (fill-mask.py in this repo)

❯ uv run fill-mask.py LFM2.5-Encoder-230M-F16.gguf "The capital of France is [MASK]."

top-5 at [MASK]:
#   1    16.17  ' Paris'
#   2    13.41  ' Strasbourg'
#   3    13.35  'Paris'
#   4    13.18  ' Lyon'
#   5    11.87  ' Versailles' 

The same server also serves per-token embeddings directly:

curl -s http://localhost:8080/embedding -d '{"content": "hello world"}'

Find more details in the original model card: https://huggingface.co/LiquidAI/LFM2.5-Encoder-230M