Skip to content

Block 3 — Summarizer model choice

Voxpost needs a local, multilingual summarizer for a worldwide product. The shipped default is csebuetnlp/mT5_multilingual_XLSum (44 languages).

Research method: web / papers first → shortlist → verify each on Hugging Face against Voxpost constraints below.

Voxpost constraints (scorecard)

# Requirement Must have
1 Multilingual — summarize in the same language as the email Yes
2 Local download — no cloud inference API Yes
3 Short output — one speakable line (~15–30 words) Yes
4 Transformers — fits existing Block 3 stack Strong preference
5 Desktop CPU — shipped desktop app, no GPU assumed Yes
6 Shippable license — Apache/MIT-style Yes
7 Maintained / proven — real downloads, model card, benchmarks Strong preference

Reality check: There is no widely adopted, multilingual, email-specific, small, local summarizer on Hugging Face. Every option is a tradeoff between language coverage, size/speed, and email domain.

What the web recommends (2024–2025)

Source / pattern Recommendation Fits Voxpost?
Multilingual summarization papers (XL-Sum, MLSUM, “Towards Unifying Multi-Lingual and Cross-Lingual Summarization”) mT5 family fine-tuned on XL-Sum or MLSUM Yes — seq2seq, many languages
Edge / low-resource NLP surveys mT5-small (~300M) or quantized mBART for CPU Partial — small mT5-XLSum forks exist but quality is poor (see below)
Email assistant projects (Vera, E.M.Pilot, local inbox tools) Small LLMs (Qwen2, Gemma, Llama) with prompts Partial — multilingual + flexible, but causal LM, slower, harder to constrain length
Email-specific HF models (wordcab, IrisWiris, 7B email LLMs) T5 on email or 7B instruct English-only or too heavy for CPU v1
Microsoft UniSumm (ACL 2023) Unified few-shot summarizer English-only on HF; weights not first-party maintained

Web consensus for “many languages + local + summarization”: mT5-on-XL-Sum, not email-tuned T5.

Hugging Face verification (candidates)

csebuetnlp/mT5_multilingual_XLSum

Field Verified on HF
Task summarization / text2text-generation
Base google/mt5-base (~580M params)
Languages 44 tagged (fr, en, ar, es, de, pt, hi, ja, zh, tr, ru, vi, …)
Downloads ~2.4M all-time — dominant multilingual summarizer on HF
Training XL-Sum (BBC news, 1.35M pairs) — ACL 2021 paper
French XL-Sum ROUGE-1 ~35.3 (published per-language table)
Output length Model config max_length=84 — good for short cues
API AutoModelForSeq2SeqLM — drop-in for Voxpost
Input Plain article text (no summarize_brief: prefix)
License Research model; check repo before commercial ship

Voxpost fit: Best match for worldwide + same-language output + Transformers. Weak on email domain (news-trained) and CPU latency (~2–5s).

Suggested Voxpost prompt shape:

From: {sender}. Subject: {subject}. {cleaned_body}

Generation: max_length=84, num_beams=4, no_repeat_ngram_size=2 (from model card).

Tier B — Viable alternatives (not default)

google/flan-t5-base (~250M)

Pros Cons
58M+ downloads; Apache 2.0; instruction-tuned Not a dedicated summarizer — needs prompt engineering
Tagged en, fr, de, ro + “multilingual” Same-language output not guaranteed
Smaller / faster than mT5-base XLSum Weaker than XLSum-finetuned models for summarization

Example prompt to test: Summarize this email in the same language as the email:\n{body}

Verdict: Reasonable fallback if XLSum is too slow, or for early prototyping — not as reliable for “always French in → French out.”

google/flan-t5-small (~80M)

Same tradeoffs as flan-t5-base, faster, lower quality.

Qwen/Qwen2.5-0.5B-Instruct (~500M)

Pros Cons
29+ languages (per Qwen2.5 paper/blog) Causal LM — new pipeline (chat template, not seq2seq)
43M+ downloads; Apache 2.0 Easy to ramble; needs strict system prompt + max_new_tokens cap
Strong for future “translate to language X” mode Slower/heavier than T5-small; CPU marginal

Verdict: Best Phase 2 candidate for translate-mode + complex instructions — not the simplest Block 3 swap.

Tier C — Keep as optional profile

wordcab/t5-small-email-summarizer (~60M)

Pros Cons
Email-trained; summarize_brief: prefix; very fast on CPU English only — confirmed failure on French mail
Apache 2.0; already integrated Wrong default for worldwide product

Verdict: Rejected — English-only; failed on French inbox mail in live tests.

Tier D — Rejected after HF check

Model HF finding Why rejected
ankitkupadhyay/mt5-small-finetuned-multilingual-xlsum ROUGE-1 ~9; 17 downloads/mo; empty model card Quality too low
T-Systems-onsite/mt5-small-sum-de-en-v2 DE + EN only Not worldwide
maan909/unisumm English only; 7 downloads/mo Not multilingual on HF
Radiantloom/radiantloom-email-assist-7b 7B LLM Too heavy for CPU v1
Walid777/llama3-8b-emails-summarization 8B Too heavy
vapit/bart-large-cnn-finetuned-for-email-and-text English BART Not multilingual
Per-language mt5-small-finetuned-xlsum-{lang} forks Single language each Doesn’t scale for “worldwide” one binary
WiseIntelligence/mT5_multilingual_XLSum-Optimum-ONNX-Quantized-AVX2 No model card; 6 downloads/mo Unmaintained; risky
GGUF XLSum builds (llama.cpp) Different runtime Out of scope until Voxpost adds GGUF path

Scoring matrix (Voxpost worldwide)

Model Multilingual Email domain CPU size Same-lang out Transformers HF trust Total
mT5_multilingual_XLSum ★★★★★ ★★☆☆☆ ★★☆☆☆ ★★★★★ ★★★★★ ★★★★★ Best default
flan-t5-base (prompted) ★★★☆☆ ★★☆☆☆ ★★★☆☆ ★★★☆☆ ★★★★★ ★★★★★ Fallback
Qwen2.5-0.5B-Instruct ★★★★☆ ★★★☆☆ ★★☆☆☆ ★★★★☆ ★★☆☆☆ ★★★★★ Phase 2 / translate
wordcab t5-small ★☆☆☆☆ ★★★★☆ ★★★★★ ★☆☆☆☆ ★★★★★ ★★★★☆ Rejected (EN-only)

Decision (updated after deep dive)

Default for worldwide product

csebuetnlp/mT5_multilingual_XLSum

Only HF model that clearly checks: 44 languages, summarization task, millions of downloads, published per-language benchmarks, short outputs, Transformers seq2seq.

Config

[summarize]
model = "csebuetnlp/mT5_multilingual_XLSum"

Optional override via VOXPOST_SUMMARIZER_MODEL for another compatible seq2seq hub id.

Pipeline mitigations (still required)

Even with mT5 XLSum:

  1. Keep email_clean.py, polish_for_tts(), is_usable_summary(), template fallback — news model on forwards/noise will still fail sometimes.
  2. Language detect → set Supertonic [tts].lang (and future [speech] mode).
  3. Optional ONNX export of XLSum later (like Block 4 Supertonic) if CPU latency blocks UX.

Implementation checklist

  • Add [summarize] to voxpost.toml (model)
  • mT5 XLSum plain-text email input adapter
  • Benchmark French forward mail + English sample on real hardware
  • Document download size ~2.3GB and RAM ~2GB+ in BLOCK_3_SUMMARIZE.md

References