Pistis: 27B and 9B Multimodal Models Trained with IDRL

It is a family of multimodal large language models at 27B and 9B parameters, built on Qwen3.6 and Qwen3.5, developed through a general and scalable post-training framework.

IDRL (Interleaved Distillation and Reinforcement Learning) is a post-training paradigm that integrates on-policy distillation and reinforcement learning within a single training loop, alternating between the two objectives rather than optimizing either in isolation or combining them in a static joint loss.

Pistis-Thinking is designed to strengthen deep multimodal reasoning, while Pistis-Agentic additionally incorporates agentic trajectory data to support long-horizon planning, iterative reasoning, and tool use, and is particularly strong in multimodal search.

The report describes research results against the corresponding base models only. It does not mention open weights, licensing, or Arabic-language performance, so adoption would require independent evaluation and compliance review.