NextTokenAI/NextSearch-1-XS
NextSearch-1-XS is a 9 billion parameter dense model developed by NextTokenAI, part of the NextSearch-1 family of web research agents. Post-trained on Qwen3.5-9B, it is designed to decompose questions, perform live web searches, reconcile evidence, and generate concise answers or structured research artifacts. Optimized for per-call accuracy, tail latency, and cost, it functions as a research component within larger orchestrator systems, serving efficiently on a single GPU.
Loading preview...
NextSearch-1-XS: A Compact Web Research Agent
NextSearch-1-XS, developed by NextTokenAI, is the smallest model in the NextSearch-1 family, specifically engineered as a web research agent. Built upon the Qwen3.5-9B base model, this 9 billion parameter dense model is designed for efficient and accurate information retrieval from the live web.
Key Capabilities
- Question Decomposition: Breaks down complex queries into manageable search tasks.
- Live Web Search & Fetching: Interacts with the live web to gather relevant information.
- Evidence Reconciliation: Processes and synthesizes potentially conflicting evidence from multiple sources.
- Concise Answering & Structured Artifacts: Generates clear, direct answers or structured research outputs.
- Optimized for System Integration: Designed to be a cost-effective and low-latency research component within larger AI orchestrator systems.
Performance Highlights
NextSearch-1-XS demonstrates strong performance on specialized web research benchmarks, often outperforming other small and cheap models. For instance, it achieves 0.289 on SEAL-0 (fresh/conflicting evidence) and 0.790 on FRAMES (multi-constraint retrieval) in live-web evaluations. When utilizing the recommended exa-auto backend, its four-bench mean score rises significantly from 0.592 to 0.693, surpassing many genuinely small models and even edging out larger anchors like nemotron-3-ultra in specific contexts. The model is released under the Apache License 2.0.
When to Use
NextSearch-1-XS is ideal for applications requiring a dedicated, efficient, and accurate web research component that can be served on a single GPU, where per-call accuracy and low operational costs are critical.