ALIA 40B Instruct 2606
ALIA 40B Instruct 2606 combines long-context instruction following with BSC's helpfulness and safety post-training.
Model and training
The card describes stronger system-prompt following, multi-turn consistency and handling of noisy requests. It uses synthetic preference data for alignment. This is the general instruction model, while the separately trained FC line handles function calling. BSC recommends low sampling temperatures and advises against a repetition penalty.ALIA 40B Instruct 2606: official model card
Practical use
Practical guidance. Compare this release on the actual documents and conversation lengths used by the application. Keep retrieval quality separate from the model's ability to use supplied evidence. If the task requires tools, evaluate the FC release directly; adding a tool schema to a chat prompt does not reproduce that model's training.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | ALIA-40b-instruct-2606ALIA-40b-instruct-2606-GGUF |
| Context window | 163,840 tokens Maximum sequence length documented for these weights; the prompt and generated text share the window. ALIA 40B Instruct 2606: official model card |
| License | Apache License 2.0.ALIA 40B Instruct 2606: official model card |
| Model size | 40B class.ALIA 40B Instruct 2606: official model card |
Related models
ALIA 40B FC 2606 adds explicit reasoning to BSC's function-calling model line.
Launch date not confirmedDownloadableALIA 40B Instruct 2605 is a distinct instruction-tuned release aimed at more precise prompt and output-format control.
Launch date not confirmedDownloadable