INTELLECT-1-Instruct
INTELLECT-1-Instruct adapts the distributed base model for following instructions. Arcee AI performed the post-training.
What the release provides
Post-training combined supervised fine-tuning, preference optimization, and model merges. Arcee used its MergeKit, EvolKit, and DistillKit tools. The training data covered general instructions, tools, code, mathematics, and multilingual examples.INTELLECT-1-Instruct official model card
How to use it well
Practical guidance. Keep the supplied tokenizer and chat format with the model. Check tool-call formatting separately from answer quality before connecting it to software.
The chat format affects the result
The release notes report much lower initial training loss with the Llama 3.1 chat template than with ChatML without an explicit start token. The card also requires a beginning-of-sequence token.INTELLECT-1-Instruct official model card
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Context window | 8,192 tokens Configured position limit. Serving limits depend on the runtime and settings. INTELLECT-1-Instruct model configuration |
| Base model developer | Prime Intellect; INTELLECT-1INTELLECT-1-Instruct official model card |
| Repository license | Apache 2.0INTELLECT-1-Instruct official model card |
| Parameters | 10 billionINTELLECT-1-Instruct official model card |
| Post-training developer | Arcee AIINTELLECT-1-Instruct official model card |
Related models
- INTELLECT-1
Prime Intellect ยท INTELLECT
A 10-billion-parameter base language model trained across distributed computers.
Launch date Downloadable