Pistis Technical Report Introduces 9B and 27B Multimodal Models

Summary

The Pistis technical report introduces a family of 27B- and 9B-parameter multimodal large language models built on Qwen3.6 and Qwen3.5, respectively. The models are developed with a post-training framework that first uses large-scale multimodal supervised fine-tuning to establish a foundation. The report then proposes Interleaved Distillation and Reinforcement Learning (IDRL), which alternates on-policy distillation and reinforcement learning in one training loop instead of using isolated objectives or a static combined loss. The authors say this design improves knowledge transfer, optimization stability, and credit assignment for long-horizon agentic trajectories while reducing capability trade-offs. At each model scale, Pistis provides Pistis-Thinking for deeper multimodal reasoning and Pistis-Agentic, which adds agentic trajectory data for planning, iterative reasoning, and tool use. Pistis-Agentic is reported to be particularly strong in multimodal search, and both variants outperform their corresponding base models. The report also introduces Pistis-Auto-Harnessing (PAH), a system-level method that iteratively optimizes the agent’s inference harness. PAH improves performance without changing model parameters or increasing the interaction budget.