MT0 Models by BigScience
Last refreshed 2026-05-19. Next refresh: weekly.
Details
About
MT0 (mT0) is BigScience’s multilingual instruction-following family made by multitask finetuning mT5 on the crosslingual xP3 mixture. Together with BLOOMZ, the project targets zero-shot prompting in dozens of languages: models follow natural-language instructions and generalize to unseen tasks and languages. The published mt0 sizes run from small through xxl (including mt0-base and mt0-large), with xP3-tuned variants recommended for English prompting and xP3mt variants when prompting in non-English.
Decision facts
Current Variants
Use-when guidance is based on each model's tracked capabilities, context window, release date, and replacement status.
Use when the workload needs 1K context and 580M parameters.
Use when the workload needs 1K context and 1.2B parameters.
Use when the workload needs 1K context and 300M parameters.
Use when the workload needs 1K context and 3.7B parameters.
Use when the workload needs 1K context and 13B parameters.
| Model | Use when | Released | Signals | Status |
|---|---|---|---|---|
| MT0 Base | Use when the workload needs 1K context and 580M parameters. | 2024-01 | 1K context580M parameters | Current |
| MT0 Large | Use when the workload needs 1K context and 1.2B parameters. | 2024-01 | 1K context1.2B parameters | Current |
| MT0 Small | Use when the workload needs 1K context and 300M parameters. | 2024-01 | 1K context300M parameters | Current |
| MT0 XL | Use when the workload needs 1K context and 3.7B parameters. | 2024-01 | 1K context3.7B parameters | Current |
| MT0 XXL | Use when the workload needs 1K context and 13B parameters. | 2024-01 | 1K context13B parameters | Current |
Release Timeline
1 release groupSpecifications(5 models)
Available From(1 provider)
Pricing
| Model | Provider | Input / 1M | Output / 1M | Type |
|---|---|---|---|---|
| MT0 XXL | IBM watsonx | $1.8 | $1.8 | Serverless |