I used Doubao to learn K12 classes with my son. And some of the responses are interesting but most of the answers are not accurate enough.

I just tried to use Coze AI Agent Builder (a similar product to OpenAI GPT store) to build a free AI Agent Chatbot using the default template and publish online. First of all, the process is quite simple and you can use UI interface to easily publish a default Chatbot without any modification and using the default LLM choices. You can also publish your GPT or LLM apps using various new LLM. And this is the address of the Chatbot I just published. Check this out: https://www.coze.cn/store/agent/7451157046348775476?bid=6epujr7no4015&bot_id=true

This AI generated picture of female movie and TV opera star Dilraba Dilmurat is great. The humanoid is realistic and look just the same as Dilraba. And the color of the eyes are correct.

Well, I have to say the picture generated by Doubao doesn't make me feel very comfortable. The reason is the earth is too large compared to the moon base and the perspective and relative size is not correct. Additionally, I think the pits on the earth look terrible to me. It's earth in the background as in the prompt, not the mars. Really Bad.

Overall, the responses from AI boyfriend of Doubao Chatbot are natural and engaging. But some responses are not related to the context of the dialogue. The last question is "I am breaking up with you". The bot respond with some general replies and seems unrelated to the context of the dialogue, which is "the boyfriends doesn't know much about taste of movies.".

The answers are very detailed with trustworthy data sources.

The face of the generated Lady gaga in the rain looks really strange and quite evil

Wrong number of arms of professor fox.

Again, Three Armed Peggy Pig Eating Bacon.

I asked Doubao to generate surreal style picture of a professor fox teaching lessons to animals. 3 arms professor, so not good enough.

Well the picture from Doubao is not satisfying. The prompt is to generate a "Melting ice cream, galaxy background, Multispectral Imaging, aliasing, glow, shaders, vivid, photorealist, hyper realist". But it seems like the ice cream and the galaxy background are from two different layers and not aligned with each other.

The promote in this AI girlfriend agents is "Cool girl not very easy to get along with" I tried to tease the bot and ask what to do on the weekends and I want to break up with her. The replies I got are somehow are consistent with the promote. But the qualities of replies I not very high. Especially is too short and I don't know what to ask next.

good reply from AI girlfriend

I asked Double who is the second NBA scorer in hisyory. The answer is correct and Jabbar and it also provides some details of sources.

I asked Qwen3 to help me with the coding problem, which is to create a package folder structure for both python and typescript. It should also contains a folder for plugin. Right now. Qwen3 provides the best answer to me compared to DeepSeek and many other.

Qwen 0.6B small size LLM is extremely powerful in realworld applications such as search and recommendation, query intent recognition, etc. And Qwen3 0.6B model is the SOTA one compared to previous counterparts such as Gemini and Llama small size LLM.

Qwen3 32B model series are the most widely adopted and deployed model in industrial applications, which compromise of inference speed and performance. This updated version of Qwen3 32B model have the thinking mode and non-thinking mode, which supports both the common task of chat/text generation and more complex task of math, code generation, etc. On the AIME and many other math benchmarks, Qwen3 surpass many of the opensource counterpart.

Qwen3 235B A22B model is more like an upgraded version of DeepSeek-R1. And it is also compared with Deepseek R1 model on multiple benchmarks of code and math. Personally, I don't Qwen3 is a huge upgrade compared to Gemini/OpenAI and Deepseek model, but more like a compromised version of complex thinking and realistic usage.

Correct result: \sqrt(3)/3. To test the geometry question on Qwen app and the thinking mode you can get the result: Thinking mode: correct answer \sqrt(3)/3. Without thinking mode: wrong answer. Overall, the 235B model is quite powerful compared to previous SOTA model. More about the key updates in Qwen3: Hybrid reasoning model, expanded language support (100+ languages), enhanced tool calling capabilities with Qwen-Agent supporting MCP. The newly open-sourced Qwen3 is China's first "hybrid reasoning model", a concept initially proposed by Claude3.7 and recently adopted by Gemini2.5 Flash. Essentially, this allows the model to toggle reasoning processes on/off. The primary purpose is to accelerate response generation for simple queries or time-sensitive scenarios by optionally disabling the thinking process while maintaining output quality. Previous approaches struggled to directly suppress reasoning steps in LLMs without retraining, as prompt engineering offered limited control. Qwen3 introduces two control methods: 1) A hard switch via enable_thinking parameter (True/False), and 2) When enabled, secondary soft switching through appending /no_think or /think tokens. Qwen also provides recommended parameter configurations to ensure optimal performance: Think mode: Temperature=0.6, TopP=0.95, TopK=20, MinP=0 Non-think mode: Temperature=0.7, TopP=0.8, TopK=20, MinP=0 Additionally, Qwen3 features specialized training for tool invocation, with Qwen-Agent now supporting MCP.

Qwen3 235B is the largest MoE model of the Qwen3 series. I tried some multi-hop.logic question on the Qwen App and the response generation is really fast. I asked what happened in the country when Elon Musk 's Mom give birth to him? The model summarized the result of war and many situation in South Africa.

The logic question from Qwen3 is correct. The relation is the divorce and current girlfriend relationship. It successfully answered Feng Qing. Besides, the speed genration of first token is really fast.

Qwen 2.5 models have already been widely adopted on many industrial applications. And Qwen 3 release will be even more exciting and I think it will set new standard of LLM and even the large reasoning models.

The first picture generated by Qwen is not Skye in paw patrol. So I have to add the word "Skye" into the prompt, so at least it's showing me a female cartoon character. The remaining is not correct.

I like the color of moon base picture generated by Qwen, very surreal and futuristic. The only drawback is that there are two planets in the backgrounds. One is the earth as requested in the prompt. The other one seems redundant.

The reponses from AI boyfriend of Qwen are very natural and engaging as well. I would very much like to chat with this robot next time.

Qwen answered the largest Penguins live on antarctic continent. Including some islands such as Coats Land near Weddell Sea.

The facial expressions are a little bit weird. Other details are great.

It looks good overall. But the face doesn't look like Trump that much

The picture of professor fox looks amazing.Detailed facial expressions.

I asked Qwen to generate an image of melting ice cream with glow, shaders, vivid color tunes. Overall, the image looks amazing and I likes the glowing stars in the backgrounds. The only drawback is that the top of the ice cream is not melt.

Overall, I think the AI girlfriend's replies are interesting. Especially when I implement our music taste is different. The prompt is that she is a rock music lover.