Pistis introduces 27B- and 9B-parameter multimodal models via new training framework
Researchers released the Pistis model family on arXiv, featuring two multimodal large language models built on Qwen3.6 and Qwen3.5. The 27B-parameter model uses Qwen3.6, while the 9B-parameter model is based on Qwen3.5. Both leverage a novel post-training framework called Interleaved Distillation and Reinforcement Learning (IDRL), which alternates between on-policy distillation and reinforcement…
Key points
- Pistis model family includes 27B-parameter (Qwen3.6-based) and 9B-parameter (Qwen3.5-based) multimodal LLMs
- IDRL framework alternates distillation and reinforcement learning for stronger performance and stability
- Pistis-Agentic variant excels in multimodal search and supports long-horizon agentic tasks
The framework produces two specialized variants per scale: Pistis-Thinking for deep multimodal reasoning and Pistis-Agentic for long-horizon planning, iterative reasoning, and tool use. Pistis-Agentic excels in multimodal search. The team also introduced Pistis-Auto-Harnessing (PAH), a system-level method that optimizes inference harnesses without altering model parameters or increasing interaction costs. Experiments show PAH improves performance without parameter updates.
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Generative AI & Models
All →- Apiiro finds 1,122 public GitHub files with secrets saved by Claude Code · 1 src
- Thinking Machines releases Inkling 2 Small, a 594B MoE model with 17B active parameters · 8 src
- OpenAI releases GPT-6 Cyber, a Critical-rated model that outperforms predecessors in vulnerability detection and alignment · 25 src
- Google releases Gemini Omni 1.2 Flash with 15-second video generation and 30% price cut to $0.07 per second · 10 src
- US accuses Chinese executive of allegedly sharing Nvidia chip info with Beijing · 11 src
Comments
via GitHub Discussions