Falcon H1 Tiny 90M Instruct
Falcon H1 Tiny 90M Instruct is a compact English language model with downloadable weights. It combines transformer attention with Mamba processing and uses the Falcon-LLM License.
An experiment in when to teach instructions
TII compared conventional pretraining followed by instruction tuning with a mixture that includes instruction data during pretraining. The release preserves base, curriculum and pre-DPO checkpoints for research. The main Instruct checkpoint is distinct from those intermediate variants.Falcon H1 Tiny instruction-training study
Start with a narrow output task
Practical guidance. A short rewrite or field extraction can reveal whether a model this small is useful for a specific application. Include requests with missing information and check that it does not invent the answer. Its size alone does not establish performance on a device.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | tiiuae/Falcon-H1-Tiny-90M-InstructFalcon-H1-Tiny-90M-BaseFalcon-H1-Tiny-90M-Instruct-Curriculum |
| Nominal size | 90M parametersFalcon H1 Tiny 90M Instruct model card |
| Available formats | Model weights and an official GGUF conversionFalcon H1 Tiny 90M Instruct model card |
Related models
A multilingual H1 Tiny language model with about 100M parameters and a 65,000-token vocabulary.
Launch date not confirmedDownloadable- Falcon H1 Tiny Tool Calling 90M
Technology Innovation Institute (TII) · Falcon H1 Tiny
Falcon H1 Tiny Tool Calling is a 90M model for selecting functions and generating their arguments.
Launch date not confirmedDownloadable