A new family of multimodal large language models, named Pistis, has been introduced in a technical report. Comprising 27B and 9B parameter scales, these models are built upon the Qwen3.6 and Qwen3.5 architectures, respectively, utilizing a general and scalable post-training framework to enhance performance over their base versions.
The IDRL Framework
Central to the Pistis development is the Interleaved Distillation and Reinforcement Learning (IDRL) paradigm. This novel approach tightly integrates on-policy distillation and reinforcement learning within a single training loop. By alternating between these objectives rather than optimizing them in isolation, IDRL facilitates more effective knowledge transfer and greater stability during the training process, particularly for long-horizon agentic trajectories.
Specialized Variants and System Optimization
The framework yields two distinct variants for each model scale:
- Pistis-Thinking: Optimized for strengthening deep multimodal reasoning capabilities.
- Pistis-Agentic: Incorporates agentic trajectory data to support complex planning, iterative reasoning, and tool utilization, showing particular strength in multimodal search.
Alongside parameter optimization, the researchers introduced Pistis-Auto-Harnessing (PAH). This system-level method automatically improves the agent's inference harness through iterative optimization. Notably, PAH enhances performance without requiring model parameter updates or increasing the interaction budget, representing a significant efficiency gain in model deployment.
