DeepSeek text captchas

Tim

Hey guys, 

deepseek stopped allowing upping balance, I didn't up it before and now it's not usable. 

I am running local copy of deepseek r1 model via LM Studio. Is it possible to provide custom api url for this option in SER? @Sven please help. I looked into SER installation directory but couldn't find anywhere to change it from official API address.

Thanks, Tim

royalmice

@Tim ,

Deepseek is still operational, but they have temporarily suspended the ability to top up balances. I believe this decision was made due to the surge in popularity of their free AI module, which led to an influx of users from the Play Store and Apple Store. Unfortunately, this also made them a target for DDoS attacks and repeated hacking attempts from Western companies. Restricting access to their API was likely one of their measures to mitigate these threats.

The Deepseek API is still functional for me since I managed to top up a sufficient balance before they halted top-ups.

I also use a local copy of Deepseek with Ollama, but I am unsure how one would integrate it with GSA.

One possible alternative could be for @Sven to add OpenRouter to the GSA CAPTCHA solver list, as Deepseek modules are available on OpenRouter. However, I am not certain which exact module would be best suited for this use. If Sven were to integrate OpenRouter into GSA CAPTCHA options, users would need to add their OpenRouter API key and have the ability to select modules that support text CAPTCHA solving. This would require some testing, but hopefully, one of the free models available on OpenRouter supports CAPTCHA solving.

This could also serve as a potential workaround for OpenAI’s chatGPT rule, which prevents CAPTCHA-solving through OpenAI's API. Since OpenRouter would be used as the intermediary, OpenAI would recognize the request as coming from OpenRouter rather than directly from an individual account, reducing the risk of bans.

Tim

I am already using seo content machine with local deepseek via lm studio. All i had to do was add api url into settings as 

http://127.0.0.1:1234/v1/chat/completions

So my thinking was if something like this can be added to GSA SER, as an option when selecting deepseek in captchas.

royalmice

Tim said:

I am already using seo content machine with local deepseek via lm studio. All i had to do was add api url into settings as 

http://127.0.0.1:1234/v1/chat/completions

So my thinking was if something like this can be added to GSA SER, as an option when selecting deepseek in captchas.

Nice. I tried creating content with Deepseek, but the quality was not so good. I prefer to use Claude Sonnet 3.5 and Zimmwriter.

I am curious which deepseek model you are running locally. I was crazy and downloaded the 70b model, but that is too much for my machine, and it takes a while to get the output; if I am going to use it for captcha solving, it will have to be 7b or 8b .

=========

I don't have the time to test it right now, but try editing the deepseek.ini file in GSA SER

Change the API URL

Locate this line in [text_solve] url=https://api.deepseek.com/chat/completions

Replace it with: url=http://127.0.0.1:1234/v1/chat/completions

Remove API Key Header (Optional)

Since your local Deepseek does not require an API key, you can either remove or comment out the following line: 

add header=Authorization: Bearer %Api-Key%

Adjust the model Name (If Needed)

The post_data section currently contains: 

 post_data={"model":"deepseek-chat","messages": [{"role":"system","content":"You are a helpful assistant."},{"role":"user","content":"%prompt%"}],"stream": false}

Ensure that "deepseek-chat" matches the model name that you have installed 

Tim

Thanks for instruction, everything connected perfectly, and indeed I am using deepseek-r1-distill-llama-8b model. 

Somehow this model doesn't answer in one or few words, no matter what instructions i give. It outputs bunch of reasoning inside <think> tags, and then in the end  of text - one word answer such as sky is blue. So when I test deepseek captcha in SER it gives very long detailed answer with line breaks etc. I am using "answer only without elaborating or explaining" prompt.

Tim

Seems like it's showing it's thought process, when I ask same question in chat window it hides thoughts and shows just word blue

royalmice

Tim said:

Seems like it's showing it's thought process, when I ask same question in chat window it hides thoughts and shows just word blue

Tim, i think we need to find the right model. Deepseek Captcha solving existed long before the R1 model came out so maybe some older model need to be used. 

royalmice

If we look at the code again : 

post_data={"model":"deepseek-chat","messages": [{"role":"system","content":"You are a helpful assistant."},{"role":"user","content":"%prompt%"}],"stream": false}

We can see the Deepseek model is : deepseek-chat
We can see : "stream": false    ==>  setting means that the response is returned as a complete message, rather than streaming partial words as they are generated. This avoids the "thinking" effect that occurs when models process and return text incrementally.

Make sure you use the same model as the api model. You mentioned you use  lm studio, i have not used that, I use Ollama, so in my case it would be the below command:
ollama list
ollama pull deepseek-chat

I will have some time on the weekend to have a go at this.

Tim

Yeah it is set to stream : false. Anyway, DeepSeek now allows top up again, so guess problem solved, i'll just use that. Thank you very much @royalmice for all the help