summarization
HF group: NLP · Status: 🟡 partial (works through chat LLM)
What it is
Long text → short text. Abstractive summarization.
Open-weight models
| Model | Params | Released | License | Quality | Notes |
|---|---|---|---|---|---|
| BART-Large-CNN | 400 M | 2020 | MIT | Foundational | Old but reliable. |
| Pegasus-XSum | 570 M | 2020 | Apache-2.0 | Strong news summarization | Older. |
| Any chat LLM | 1.5-30 B | 2024-26 | various | Best | LLM with a “summarize” prompt outperforms specialized models. |
Infrastructure required
Inference
- 🟡 Works via chat LLM today.
- ❌ Specialist encoder-decoder (BART/Pegasus) — same gap as
translation.
Input
- Plain text.
Output
- Plain text (typically short).
Storage
- ✅ Weights cache.
Interaction (IPC + SDK)
- 🟡 Today:
llm.chatwith a “summarize” prompt. - App pattern: long-doc summarization typically chunks input first.
Capabilities (manifest)
capabilities.models[]for chat LLM (or specialist).
Gaps
Nothing critical. Specialist summarization models matter if
deterministic output structure is required, in which case they
share the encoder-decoder runtime gap with
translation.
See also
text-to-texttranslation— same encoder-decoder need- Index:
../modalities-and-models-survey.md