Falcon H1 Tiny R 90M
Falcon H1 Tiny R 90M is a downloadable reasoning model with a hybrid transformer and Mamba architecture. It is part of TII's experiment in training small models directly on reasoning data.
How this checkpoint was trained
The 90M model follows the reasoning-only pretraining approach. TII reports that it is more prone to repeated text than the larger Tiny R model. The report does not claim that the 90M checkpoint received the same GRPO stage as the 0.6B release.Falcon H1 Tiny reasoning study
Use a fixed reasoning budget
Practical guidance. Include simple problems with exact answers and requests that require several steps. Check whether the smaller model finishes the task within the budget. Repetition penalties can alter output, so retain the settings with each evaluation result.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | tiiuae/Falcon-H1-Tiny-R-90M |
| License | Falcon-LLM LicenseFalcon H1 Tiny R 90M model card |
| Nominal size | 90M parametersFalcon H1 Tiny R 90M model card |
Related models
- Falcon H1 Tiny R 0.6B
Technology Innovation Institute (TII) · Falcon H1 Tiny
Falcon H1 Tiny R 0.6B is a downloadable reasoning model with a hybrid transformer and Mamba architecture.
Launch date not confirmedDownloadable - Falcon H1R 7B
Technology Innovation Institute (TII) · Falcon H1R
Falcon H1R 7B is a reasoning-focused adaptation of Falcon-H1-7B-Base.
Launch date not confirmedDownloadable