Falcon H1 Tiny Multilingual 100M

Falcon H1 Tiny Multilingual expands the English Tiny design to multilingual text. TII increased the vocabulary to about 65,000 tokens, which raises the model size to about 100M parameters.

Vocabulary size is a real design change

The study uses multilingual web content, Wikipedia and textbooks. It also compares separate fine-tuning with instruction data mixed into pretraining. The repository name and technical study identify the multilingual 100M release; some inherited card fields still describe the English 90M template.Falcon H1 Tiny multilingual study

Check that the output stays in the requested language

Practical guidance. Give the model the same task in each required language. Inspect language switching, names and numbers as separate failure types. A multilingual vocabulary helps represent text, but it does not establish equal capability across languages.

Specifications & access

Falcon H1 Tiny Multilingual 100M specifications
DetailValue and conditions
AccessDownloadable
Model identifierstiiuae/Falcon-H1-Tiny-Multilingual-100M-InstructFalcon-H1-Tiny-Multilingual-100M-Base
Architecture24 layers with a hidden width of 512Falcon H1 Tiny multilingual study
VariantsBase and Instruct weightsFalcon H1 Tiny multilingual study
LicenseFalcon-LLM LicenseFalcon H1 Tiny Multilingual 100M model card

Related models