Falcon H1 0.5B Instruct
Falcon H1 0.5B Instruct is the smallest original H1 instruction model. It uses hybrid transformer and Mamba processing for English text.
A small model with a shorter context
This checkpoint's configuration sets 16,384 positions, 36 layers and a hidden width of 1,024. The 256K family headline does not apply to this instruction checkpoint.Falcon-H1-0.5B-Instruct configuration
Use a constrained response format
Practical guidance. Try short extraction and rewriting tasks before adding long conversation history. Measure whether this model can follow the required output format with few examples. Increasing a server's context setting does not establish accuracy outside the checkpoint's documented range.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | tiiuae/Falcon-H1-0.5B-InstructFalcon-H1-0.5B-Base |
| Context window | 16,384 tokens max_position_embeddings in the instruction checkpoint configuration; effective recall depends on the task. Falcon-H1-0.5B-Instruct configuration |
| License | Falcon-LLM LicenseFalcon-H1-0.5B-Instruct model card |
| Layers and width | 36 layers; hidden width 1024Falcon-H1-0.5B-Instruct configuration |
Related models
Falcon H1 7B Instruct is a multilingual hybrid model with a published 262,144-position configuration.
Launch date not confirmedDownloadable