Alibaba PAI released SearchQwen2.5-3B, a 3B-parameter search agent model
Alibaba Cloud PAI released SearchQwen2.5-3B, a 3.09B-parameter search agent model based on Qwen2.5-3B-Instruct, on Hugging Face. It is trained with EasyDistill 2.0 and SynSearch-Data, supports structured tool calls, and has a 32,768 context length.
China context
- Original name
- SearchQwen2.5-3B
- 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 search agent pipelines, but must provide their own search/browse backend.
- For investors
- Alibaba PAI is actively releasing small, specialized agent models, indicating a strategy to capture developer mindshare in the search agent space.
Translated from Chinese. Quotes and facts link to the original sources.
Alibaba Cloud PAI released SearchQwen2.5-3B, a compact search agent model based on Qwen2.5-3B-Instruct, on Hugging Face. The model is trained with environment-aligned, solver-verified search trajectories generated by EasyDistill 2.0 and SynSearch-Data. It supports structured search and browse tool calls, with a context length of 32,768 tokens. Reported LLM-judge accuracy averages show improvements over the base model in multi-hop QA and deep search tasks.
The model uses a structured tool-call interface and requires an external search/browse backend. Reported LLM-judge accuracy: Tool-Call SearchQwen2.5-3B achieves 48.58 on multi-hop QA and 21.40 on deep search, versus 36.10 and 7.05 for the base Qwen2.5-3B-Instruct under the same interface. These are company-reported figures.
Developers outside China can now access a small, open-weights search agent model that reduces the cost of building multi-hop search and evidence-integration applications compared to larger models. Alibaba PAI's release puts a 3B-parameter agent model on Hugging Face, directly competing with other open search agents for low-resource deployment.
For enterprises needing on-premise or low-latency search agents, this model offers a smaller footprint than larger Qwen models, potentially lowering inference costs. However, an external search backend is required, so total cost of ownership includes that infrastructure.
The next verifiable signal is whether Alibaba PAI releases the SynSearch-Data training dataset or EasyDistill 2.0 framework publicly, and whether independent benchmarks confirm the reported accuracy gains. Adoption can be checked via Hugging Face downloads and community fine-tunes.