idrak is SJM Labs' in-house bilingual language model family, built for Arabic and English from the ground up rather than adapted after the fact.
The idrak family includes two variants. idrak-1, a full-scale fine-tuned model built for depth and reasoning, and idrak-mini, a lightweight variant optimized for fast, low-cost inference.
Both models were fine-tuned end to end on a custom-curated instruction dataset of more than 230,000 examples, sourced, cleaned, deduplicated, and quality-calibrated in-house rather than pulled from an off-the-shelf corpus.
idrak reflects the same philosophy behind everything SJM Labs builds. We do not wrap someone else's API and call it AI. We train the model, own the weights, and control the entire pipeline from raw data to deployment.
idrak is currently in review and will be available soon.