ScienceOne-AI/S1-DeepResearch-32B
ScienceOne-AI/S1-DeepResearch-32B is a 32 billion parameter end-to-end agentic model developed by ScienceOne AI, designed for long-horizon deep research tasks. It features a 128K context window and excels in long-chain complex reasoning, deep research instruction following, report writing, file understanding/generation, and skills utilization. The model significantly outperforms its base Qwen3-32B on 20 agentic capability benchmarks, demonstrating performance comparable to mainstream closed-source flagship models.
Loading preview...
S1-DeepResearch-32B: End-to-End Deep Research Agent
S1-DeepResearch-32B, developed by ScienceOne AI, is a 32 billion parameter model specifically engineered for long-horizon deep research. It goes beyond typical LLM capabilities by focusing on real-world deployment in complex research scenarios. The model is built upon a Qwen3-32B base and features an ultra-long 128K context window, enabling it to handle extensive evidence chains and multi-turn interactions.
Key Capabilities
- Long-chain complex reasoning: Supports sustained, multi-stage reasoning, integrating information across documents, managing state memory, and iterating policies to achieve stable and reliable conclusions.
- Deep research instruction following: Parses multi-constraint instructions, building an understanding paradigm from task definition to result presentation, ensuring controllable processes and aligned outputs.
- Deep research report writing: Generates arguable, citable reports by organizing multi-source material and evidence checks, balancing structure, readability, and traceability for scientific writing.
- File understanding and generation: Processes various modalities like PDFs, tables, and web pages for input, and produces structured, deliverable outputs, closing the loop of parse, process, and generate.
- Skills Using: Organizes and dynamically assembles callable modules for tasks such as literature search, data analysis, experiment design, and visualization, supporting continuous research workflows.
- Long-horizon tool calling: Stably runs over 150 consecutive tool-call rounds, enabling continuous planning, execution, and self-correction with a native tool ecosystem including search, web browsing, and code execution.
Performance
Evaluated across 20 agentic capability benchmarks grouped into five dimensions, S1-DeepResearch-32B shows a clear advantage over its base Qwen3-32B and even larger Qwen3-235B models. Its overall performance is reported to be close to mainstream closed-source flagship models like GPT 5.2, Claude 4.6, and GLM-5, indicating readiness for real-world business deployment.