Skip to content
← Models
Open weightsBigScience/10 trustVerified

mT0

Multilingual T5 instruction-tuned on the crosslingual xP3 mixture, following human instructions in dozens of languages zero-shot.

Visit official page →

About

mT0 is BigScience's answer to multilingual instruction following. It takes the mT5 pretrained checkpoints, from 300M up to 13B parameters, and fine-tunes them on xP3, a crosslingual task mixture spanning dozens of languages. The result is a family of text-to-text models that can follow natural-language instructions, translate, summarize, and answer questions in languages they were never explicitly trained to handle, a capability the authors call crosslingual generalization.

Released alongside BLOOMZ in late 2022, mT0 became a standard multilingual baseline for research and a popular starting point for fine-tuning. The weights are Apache-2.0 licensed and available in five sizes from mt0-small to mt0-xxl, plus an mt variant tuned for non-English prompting. Its context window is the T5-standard 512 tokens, which limits long-document use.

Specs

Context window

512 tokens

Pricing

Free (open weights, Apache-2.0)

Free quota

Free — self-hosted weights; Hugging Face Inference API

API access

No official API; Hugging Face weights + Inference API

Weights

Open weights

Released

Not disclosed

Benchmarks

No published BLEU/COMET-style MT benchmarks; paper evaluates zero-shot crosslingual task generalization across held-out tasks and languages (author-reported)

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

multilingual-nlpmachine-translationinstruction-followingcrosslingual-qasummarization

Trust score — 8/10

/10 trustVerified

Scored 8/10 (verified tier, metrics-v1). Evidence: BigScience workshop release with the crosslingual generalization paper confirmed via the Hugging Face model cards; widely cited as a multilingual instruction-tuning baseline. Strengths: permissive Apache-2.0 license, five size options, strong zero-shot crosslingual generalization, huge research adoption. Docked for: (1) 512-token context limits document work; (2) superseded by newer multilingual instruction models for raw quality. License: Apache-2.0.

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
9.0/10
Maintenance · recency, releases, issue load
6.0/10
Openness · license + docs
10.0/10
Manual review · sanity read + red-flag check
8.0/10
  • · BigScience workshop release (Oct 2022) with crosslingual generalization paper
  • · Five sizes 300M-13B, standard multilingual instruction-tuning baseline
  • · 512-token context limits long-document use

Sources