SearchAgent-8B: Tencent’s Open-Source AI for Deep Research

Social media is full of AI slop, but sometimes you find a real gem like search agent 8 billion. This is an open-source AI model designed to do deep research. Standard chatbots just guess an answer, this model acts like a detective. If you ask it a complex question, it doesn't just answer. It breaks the problem down, performs 10 to 15 separate research queries against a local database, gathers the evidence, verifies the information, and then gives you a final answer.

Before I install and run this model, the most important point or the important feature is that usually training a model to use tools is incredibly expensive because you have to pay for every search query using APIs like serp or Google.

This team from Tencent China gives you an immense insight how these Chinese labs are trying to generate the models but at the same time due to the budget constraints due to the sanctions they are trying to save money as much as possible. They realized they could cut that cost of training to zero by not using the live web at all.

What Tencents SearchAgent-8B Is Built On

This model is based on Qwen 3. There are other three variants too. One is 30 billion and one is without HQ. The HQ model is based on 38 billion and trained with outlier suppression training strategy which I will explain shortly. This is without HQ and again it is also trained without outlier expression training strategy. And then we have this search agent A3B. Again this is the model based on coin 330 billion A3B mixture of expert model trained without outlier expression training strategy.

Outlier Suppression Training Strategy

This is just to improve the quality. It is a technique designed to stabilize reinforcement learning by aggressively terminating abnormal trajectories that show extreme behavior such as generating a burst of parallel tool calls, repeating search queries or producing malformed tool syntax.

Instead of applying partial penalties or ignoring these error, this strategy immediately stops the generation process and assigns a reward of zero.

That prevents these high variance samples from polluting the updates and significantly reduces training noise. All in all it improves the quality of the model, and we might see this innovation being used in various models in 2026.

Installing and Serving Tencents SearchAgent-8B With vLLM

My system is Ubuntu and I have one GPU card Nvidia RTX 6000 with 48 GB of VRAM.

Steps

• Install vLLM.
• Download the model and serve it locally with vLLM.
• Wait a couple of minutes for the model to load.
• Keep the server running and confirm it is serving on localhost.

While serving, there are four shards of it and it is now being served on the local system. It takes around 15.2 GB of VRAM.

Using Tencents SearchAgent-8B On a Private Corpus

For the purpose of using this model I access the model at localhost. I have defined a tool which is going to do the search on the basis of the search query which this model is going to generate, and this is quite flexible. You can then make that search to any local private offline corpora.

Example: Querying a PDF

I have this PDF file on my local system. In the risk management section, right in the middle of the document, there is a line that says: cyber security investment of $67 million enhanced our defense capabilities and then there were no cyber security incident and all that stuff.

My question was: did we have any security breaches lasted and how much did we spend on defense?

What Happened

• The model made a tool call on the basis of the natural language prompt.
• It generated the tool call according to the tool I defined.
• The tool executed, searched the corpus, and found the answer and the context.
• It found a needle in the haystack and then gave me the final answer.
• The budget figure was returned with a citation according to the report, all verifiable. After thinking it has given us the response.

Pretty targeted and directed, and that is what really really good about it, that it can do a targeted search on such a cheap budget which is the main attraction for me. You can even go with the 30 billion one.

Final Thoughts on Tencents SearchAgent-8B

I am very pumped about the next year. I think we are going to see lot and lot of innovation around cost optim optimization, quality enhancement, and also the models and the AI tooling which is going to enhance business value because that is what we needed. We really don't need AGI and all that mumbo jumbo which is going on social media. This is what matters for the bottom line of companies.