ScienceOne-AI/S1-DeepResearch-32B

S1-DeepResearch-32B: An Agentic Model for Deep Research

S1-DeepResearch-32B, developed by ScienceOne-AI, is an end-to-end agentic model specifically designed for long-horizon deep research. It emphasizes real-world deployment by integrating advanced capabilities beyond typical long-chain complex reasoning.

Key Capabilities

• Long-chain complex reasoning: Supports multi-stage, multi-hop tasks through cross-document retrieval, evidence aggregation, and policy iteration, ensuring stable reasoning and reliable conclusions.
• Deep research instruction following: Parses multi-constraint instructions across the full research chain, from task definition to tool execution and result presentation, maintaining control and predictability.
• Deep research report writing: Generates arguable, citable reports by integrating multi-source material and evidence checks, suitable for scientific writing.
• File understanding and generation: Processes various modalities (PDFs, tables, web pages) for input and produces structured outputs, closing the loop of parse, process, and generate.
• Skills Using: Organizes and dynamically assembles callable modules for tasks like literature search, data analysis, experiment design, and report generation.

Unique Features & Performance

This model features an ultra-long 128K context window for extended evidence chains and multi-turn interactions. It supports long-horizon tool calling, stably running over 150 consecutive tool-call rounds with 9 built-in common tools (e.g., search, web browsing, code execution). S1-DeepResearch-32B significantly outperforms its base model Qwen3-32B across 20 agentic capability benchmarks and achieves performance comparable to mainstream closed-source flagship models like GPT 5.2 and Claude 4.6, indicating its readiness for business deployment.