New LLM family Pistis enhances multimodal reasoning and agentic capabilities

2026-09-25

Researchers introduce the Pistis model family, featuring 27B and 9B parameter multimodal large language models. The models are built on Qwen3.6 and Qwen3.5 and utilize a novel post-training framework called Interleaved Distillation and Reinforcement Learning (IDRL).

VERA Brief

AI-generated. Grounded in the article and its cited sources.

Researchers have introduced the Pistis model family, which includes multimodal large language models with 27B and 9B parameters. These models are built on Qwen3.6 and Qwen3.5 and use a new post-training framework called Interleaved Distillation and Reinforcement Learning (IDRL) to enhance reasoning and agentic capabilities.

Key facts

The Pistis model family consists of multimodal large language models with 27 billion and 9 billion parameters.
These models are built upon Qwen3.6 and Qwen3.5.
A novel post-training framework called Interleaved Distillation and Reinforcement Learning (IDRL) was utilized.
The framework produces specialized variants: Pistis-Thinking for multimodal reasoning and Pistis-Agentic for planning and tool use.
Pistis-Auto-Harnessing (PAH) is a system-level method introduced in the report.

Source: arXiv · cs.AI

Reported by VERA Newswire.