You are viewing a single comment's thread from:

RE: I Am #AliveAndThriving Today! || #IAmAliveChallenge || Daily Off-Grid Jungle Journal Entry: Day #1421 - Preparing to Run and Test Qwen3.6 35B A3B on 8GB of VRAM and 64GB RAM - Lower Puna, Far East Big Island, Hawai'i - Tuesday, June 9, 2026

in We Are Alive Tribe13 days ago

Great post as always!

I have a random thought; it's probably nothing.

Qwen can do more for less cost in many cases, but I worry about security and doing anything to give an authoritative state an edge in the AI competition. Any thoughts on my random thought?

Some research I did:

...this adoption raises security concerns: the U.S. National Institute of Standards and Technology (NIST) found that DeepSeek’s open models are more susceptible to cyber risks than comparable U.S. models. Additionally, Chinese models may censor topics to comply with political pressures and domestic regulations, and using models hosted by Chinese entities could present data privacy risks.

Even if Chinese models are nearly as good as top U.S. models for much lower costs, enterprise customers developing AI-base applications may prefer U.S. models for political and security reasons.

U.S. researchers and companies increasingly rely on Chinese base models, creating long-term dependency on infrastructure with embedded censorship and potential security vulnerabilities.

Source: https://www.uscc.gov/sites/default/files/2026-03/Two_Loops--How_Chinas_Open_AI_Strategy_Reinforces_Its_Industrial_Dominance.pdf

I may be overly cautious but I don't want to be doing anything to increase China's dominance in anything, especially AI and robots. But that's just me! And I'm sure I own lots and lots of things Made In China so I'm not that concerned after all, maybe! :D

Happy middle of the week! Hang in there, the weekend is coming. :)

!ALIVE !BBH !UNI !PIZZA !LADY

Sort:  

Thank you very much, Kenny, I appreciate it! I hadn't heard about specifically about Chinese models being more susceptible to cyber risks, but I'll look into it a bit more closely. The good thing about open-weight models, which is equivalent to software's open source, is that there's nothing hidden, and they can be modified, such as with various uncensored iterations. Also, China is already way ahead in AI development. Chinese people make up a huge portion of the AI-development community, so it's not even just at the governmental level. One reason that Chinese models are used so extensively is because they are generally better and more capable. Google's Gemma 4 line is very good, and I use that too, but for the most part American-made models can't compare to Chinese models in capabilities and performance, at least not yet. Regarding the exposure of sensitive data, that would be an issue with any cloud-based model that has access to a device, such as through an agent, because anything it sees and processes goes through, and is recorded on the hosting servers. Aside from not being willing to pay to use AI, that is my main reason for wanting to run everything locally, under my own control. 😁🙏💚✨🤙

!ALIVE
!BBH
!LADY
!PIZZA
!UNICOIN

That AI model is used by #BraveBrowser and its "sister" project #BraveSearch. 🤔🧘‍♂️ The company explained that such model is efficient. 🤓😅

!LOLZ

Are you talking about Gemma 4? I haven't looked into the model or models that they're using yet. 😁🙏💚✨🤙

!LOLZ

No, I was talking about #Qwen !INDEED. 🤔🤯 The #Brave team says that such model is still accurate despite its relatively smaller size than the other popular LLMs. 🤓😏

!WEIRD
!HOPE
!WINEX

Ah, OK, I see. Yes, for its size, Qwen3.6 punches way above its weight class, depending on the specific size and quantization. I can't wait to finally get Qwen running successfully. The larger Gemma 4 is running fine, though loading Qwen still has issues. 😁🙏💚✨🤙

!ALIVE
!BBH
!HOPR
!INDEED
!PIXY

Let's !HOPE that such LLM has less security risks over time !INDEED! 🤗😊 One way for a user to reduce such risks is to make the chatbot completely offline, though it would reduce the effectiveness of the model. 🤔🤯🤓

!WEIRD
!WINEX

Well, I was finally successful in getting Qwen3.6 35B-A3B APEX IQuality to run using both ik_llama.cpp and llama-cpp-turboquant. Now I just need to tweak and refine the flags a bit more. A totally offline model is not feasible for me, as it needs to be able to do web searches and synthesis, and to connect and transact within multiple blockchain ecosystems, Hive included. 😁🙏💚✨🤙

!ALIVE
!BBH
!PIZZA
!PIXY
!UNICOIN

For those who prioritize security, such LLMs should be used on a separate computer (which may be expensive) or in a virtual machine (which will reduce its efficiency) !INDEED. 🤔🤯🤓

!WEIRD
!HOPE
!PIZZA

Brilliant answer, thanks!

I knew you would have something important to say on this, as you think carefully about everything that you do. Thanks again for the interaction, I appreciate you, my friend!

!ALIVE !BBH !UNI !PIZZA !LADY

I'm grateful that you appreciated it, and of course, you're always very welcome! The appreciation is very mutual, Kenny! 😁🙏💚✨🤙

!ALIVE
!BBH
!LADY
!PIXY
!UNICOIN


View or trade LOH tokens.


@kenny-crane, you successfully shared 0.1000 LOH with @tydynrain and you earned 0.1000 LOH as tips. (2/11 calls)

Use !LADY command to share LOH! More details available in this post.