Falcon H1 1.5B-Deep Instruct
Falcon H1 1.5B Deep Instruct uses more layers and a narrower hidden representation than standard H1 1.5B. It is a distinct multilingual checkpoint, rather than a quantized edition.
Why the Deep name matters
The configuration specifies 66 layers, a hidden width of 1,280 and 131,072 positions. TII trained the standard and Deep 1.5B models with identical settings; depth and width were the key difference.Falcon-H1-1.5B-Deep-Instruct configurationFalcon H1 architecture and release
Check the cost of depth
Practical guidance. This is a useful comparison when parameter storage is the main constraint. Measure complete task latency on the target hardware. A deeper model may change the tradeoff between answer quality and response time even when nominal size is similar.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | tiiuae/Falcon-H1-1.5B-Deep-InstructFalcon-H1-1.5B-Deep-Base |
| Context window | 131,072 tokens max_position_embeddings in the instruction checkpoint configuration; effective recall depends on the task. Falcon-H1-1.5B-Deep-Instruct configuration |
| License | Falcon-LLM LicenseFalcon-H1-1.5B-Deep-Instruct model card |
| Layers and width | 66 layers; hidden width 1280Falcon-H1-1.5B-Deep-Instruct configuration |
Related models
Falcon H1 1.5B Instruct is a multilingual hybrid model with downloadable weights.
Launch date not confirmedDownloadable