Pistis family merges distillation and reinforcement, a new architecture for training multimodal agents

A technical report published by twenty researchers on the Arkhive platform presents a new family of multimodal models called “Pistis”, developed in two sizes: the first contains 27 billion parameters based on the Qwen 3.6 model, and the second contains 9 billion parameters based on Qwen 3.5. These models rely on a general, scalable downstream training framework that begins by establishing a solid foundation through supervised fine-tuning on large-scale multimodal data, before moving on to address more complex training challenges.

The core innovation in this framework lies in integrating distillation and reinforcement learning within a unified training loop.It is the methodology that the researchers have named “Iterative Distillation and Reinforcement Learning” (IDRL). This approach relies on a regular alternation between the objectives of direct policy distillation and reinforcement learning, rather than isolating each objective in a separate stage or merging them into a fixed joint loss function. This alternation enables more efficient knowledge transfer, provides greater stability in mathematical optimization, and offers precise attribution of responsibility when executing long-horizon agent trajectories, thereby reducing mutual regressions in concurrent model capabilities.

This training pathway yields two specialized variants for each size: the first, “Pistis Thinking”, is intended to enhance deep logical reasoning in multimodal contexts; the second, “Pistis Agentic”, incorporates additional data on independent agent trajectories to support long-term planning, iterative reasoning, and tool use, while demonstrating exceptional strength in multimodal search tasks. Evaluation results have shown that both sizes, in both variants, outperform the base models from which they were derived.

The research team went beyond merely adjusting model weights by introducing a new system-level architecture they call “Automatic Packaging” (PAH).The method improves the inference environment for the intelligent agent through an iterative optimization path that is completely independent of the model’s parameters; experiments have shown that this approach raises performance efficiency without altering the stored weights and without adding any cost to interaction budgets or computational consumption during field deployment.

In our follow-up of this line of work, we see that this shift in downstream training architecture directly affects development teams and organizations in the Gulf, Egypt and the Levant that aim to build intelligent agents for extended tasks. If you lead a technical team working on process automation or building multimodal search engines, moving to open, specialized models of 9 billion or 27 billion parameters provides a genuine alternative to costly closed-source giant models. More importantly, from an engineering decision perspective, achieving stable agent performance is no longer limited to re-training weights at massive computational expense; tangible gains can now be realized by tuning inference environments and systematically alternating between reinforcement learning and distillation.