> ## Documentation Index
> Fetch the complete documentation index at: https://docs.liquid.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# LFM2.5-Encoder-350M

> 350M multilingual bidirectional encoder for fine-tuning into classification, token classification, and retrieval tasks

<a href="/lfm/models/liquid-nanos" className="back-button">← Back to Liquid Nanos</a>

LFM2.5-Encoder-350M is a multilingual bidirectional encoder built on the LFM2 architecture. At 350M parameters, it is the larger sibling in the encoder family, built for maximum downstream quality when fine-tuned into task-specific models (classification, token classification, retrieval, reranking, and semantic similarity) across 15 languages.

<div style={{display: 'flex', gap: '0.5rem', margin: '0.5rem 0 1.5rem 0'}}>
  <a href="https://huggingface.co/LiquidAI/LFM2.5-Encoder-350M" style={{padding: '0.35rem 0.7rem', borderRadius: '4px', fontSize: '0.85rem', fontWeight: 600, textDecoration: 'none', backgroundColor: '#fbbf24'}}><span style={{color: '#000'}}>HF</span></a>
</div>

<Note>
  Use LFM2.5-Encoder-350M when downstream quality matters most.
  For tighter latency and memory budgets or higher throughput, use [LFM2.5-Encoder-230M](/lfm/models/lfm25-encoder-230m).
</Note>

## Specifications

| Property            | Value                                                                                                                                     |
| ------------------- | ----------------------------------------------------------------------------------------------------------------------------------------- |
| Parameters          | \~354M                                                                                                                                    |
| Type                | Bidirectional encoder                                                                                                                     |
| Hidden Size         | 1024                                                                                                                                      |
| Vocabulary Size     | 65,536                                                                                                                                    |
| Context Length      | 8,192 tokens                                                                                                                              |
| Supported Languages | English, German, Spanish, French, Italian, Dutch, Polish, Portuguese, Arabic, Hindi, Japanese, Russian, Turkish, Vietnamese, Chinese (15) |

<div className="use-cases">
  <CardGroup cols={3}>
    <Card title="Text Classification" icon="tags">
      Sentiment, topic, intent/routing, moderation, and business-text linting.
    </Card>

    <Card title="Token Classification" icon="brackets-curly">
      Named-entity recognition, span extraction, and sequence labeling.
    </Card>

    <Card title="Fine-Tuning Backbone" icon="sliders">
      A base for dense embedding or late-interaction retrievers, NLI, and semantic similarity.
    </Card>
  </CardGroup>
</div>

## Quick Start

<Note>
  This is a base encoder that produces general-purpose representations.
  For task outputs, fine-tune it with a task head for your use case. See the [fine-tuning overview](/lfm/fine-tuning/overview) to get started.
</Note>

<Tabs>
  <Tab title="Transformers">
    **Install:**

    ```bash theme={"theme":{"light":"github-light","dark":"github-dark"}}
    pip install -U transformers
    ```

    **Run masked-token prediction:**

    ```python theme={"theme":{"light":"github-light","dark":"github-dark"}}
    from transformers import AutoModelForMaskedLM, AutoTokenizer
    import torch

    tok = AutoTokenizer.from_pretrained("LiquidAI/LFM2.5-Encoder-350M", trust_remote_code=True)
    mlm = AutoModelForMaskedLM.from_pretrained("LiquidAI/LFM2.5-Encoder-350M", trust_remote_code=True)

    text = f"The capital of France is {tok.mask_token}."
    enc = tok(text, return_tensors="pt")
    with torch.no_grad():
        logits = mlm(**enc).logits
    pos = (enc["input_ids"][0] == tok.mask_token_id).nonzero()[0].item()
    print([tok.decode([t]).strip() for t in logits[0, pos].topk(5).indices.tolist()])
    # -> ['Paris', 'Strasbourg', 'Paris', 'Lyon', 'Versailles']
    ```

    **Load the encoder body for downstream tasks:**

    For classification, token classification, regression, or retrieval, load the encoder body and attach your own head:

    ```python theme={"theme":{"light":"github-light","dark":"github-dark"}}
    from transformers import AutoModel

    body = AutoModel.from_pretrained("LiquidAI/LFM2.5-Encoder-350M", trust_remote_code=True)
    ```
  </Tab>
</Tabs>