Alibaba PAI released SearchQwen3-8B, an open-weight search agent model
Alibaba Cloud PAI released SearchQwen3-8B, an 8.19B-parameter search agent model based on Qwen3-8B, on Hugging Face under Apache-2.0. It is trained with EasyDistill 2.0 and supports structured tool calls for multi-hop search and browsing.
China context
- Original name
- SearchQwen3-8B
- Outside China
- Open weights · huggingface.co
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers outside China can download the weights from Hugging Face and integrate the model into agent frameworks that support structured tool calls, using an external search backend.
- For investors
- The release of an open-weight search agent model by Alibaba PAI may accelerate adoption of self-hosted agent solutions, potentially affecting the market for proprietary search APIs.
Translated from Chinese. Quotes and facts link to the original sources.
Alibaba Cloud PAI released SearchQwen3-8B, a compact search agent model with 8.19B parameters and 40,960 context length, on Hugging Face under Apache-2.0. The model is trained using environment-aligned, solver-verified search trajectories generated by EasyDistill 2.0, and is designed for multi-hop search, browsing, and evidence integration. It supports structured search and browse tool calls, and requires an external search/browse backend. Reported LLM-judge accuracy: Tool-Call Qwen3-8B 24.50 vs SearchQwen3-8B 35.42 on Multi-hop QA; Deep Search overall 40.31 vs 50.31.
SearchQwen3-8B is a fine-tuned version of Qwen3-8B that outputs structured tool calls for search and browse actions, enabling multi-step agentic workflows. It was distilled using EasyDistill 2.0 with solver-verified trajectories, which may improve tool-use reliability. The model card reports higher LLM-judge accuracy than the base Qwen3-8B on multi-hop QA and deep search benchmarks, but these are company-reported and not independently verified.
Developers outside China can now use an open-weight, Apache-2.0 licensed search agent model from Alibaba PAI, reducing the cost and complexity of building multi-hop search and browsing agents compared to proprietary APIs. This puts pressure on closed search-agent offerings by providing a customizable, self-hostable alternative.
Enterprises can deploy SearchQwen3-8B internally for research, customer support, or knowledge management tasks that require multi-step web search and evidence synthesis, without per-token API costs. The Apache-2.0 license allows commercial use and modification, subject to the base model's terms.
A key signal to watch is whether independent evaluations confirm the reported accuracy gains on benchmarks like GAIA and HotpotQA. Adoption can be tracked via Hugging Face downloads and community fine-tunes. The release of SynSearch-Data and EasyDistill 2.0 may lead to further distilled agent models from other teams.