Chinese open-weight AI models offer high capability at significantly lower costs than U.S. frontier systems, but face monetization and compute challenges, experts said at a Rest of World event in San Francisco.
Silicon Valley spent billions building AI tollbooths while Chinese labs made models people can actually download and run. Then Washington restricted chips, and China got better at doing more with less. It’s almost poetic: America tried to corner the future of AI and accidentally made efficiency China’s competitive advantage.
It’s worse than that even… frontier models now have guardrails restrictive enough that I can’t use them for legitimate work. And the ones that DO work have token prices at exorbitant rates, and chew through them like there’s no tomorrow.
Meanwhile, I can download Qwen3 and run it on my own hardware, and it’s about as good as the last generation US model without the annoying guardrails.
LLM weights aren’t code; the engines that run them are completely separate, so they don’t have traditional “backdoors.”
It’s possible some elaborate prompt trigger was trained into the weights, but there’s never been any evidence of that, for years. I’d say worrying about that is paranoid unless you work at the DoD or something, and some full-weights finetuning would scramble it anyway.
It’s a mixed bag TBH. The issue is that Qwen 3.8 27B is so good.
It depends on the quant you are running though, and the way you are running it. If you have 16GB or less VRAM and a generous amount of CPU RAM, and you are running a regular GGUF instead of something exotic, Qwen next is probably better, with the right hybrid inference engine.
Silicon Valley spent billions building AI tollbooths while Chinese labs made models people can actually download and run. Then Washington restricted chips, and China got better at doing more with less. It’s almost poetic: America tried to corner the future of AI and accidentally made efficiency China’s competitive advantage.
It’s worse than that even… frontier models now have guardrails restrictive enough that I can’t use them for legitimate work. And the ones that DO work have token prices at exorbitant rates, and chew through them like there’s no tomorrow.
Meanwhile, I can download Qwen3 and run it on my own hardware, and it’s about as good as the last generation US model without the annoying guardrails.
Hmm… which model am I going to choose to use?
I don’t know, but how many backdoors do each model have?
LLM weights aren’t code; the engines that run them are completely separate, so they don’t have traditional “backdoors.”
It’s possible some elaborate prompt trigger was trained into the weights, but there’s never been any evidence of that, for years. I’d say worrying about that is paranoid unless you work at the DoD or something, and some full-weights finetuning would scramble it anyway.
An infinite number of. With models, the concept of backdoors becomes a bit fuzzy.
Did you mean qwen 3.8 max? Or qwen 3 coder next? They are very different.
I’ve been rocking Qwen 3.8 27b for programming, anyone have an opinion if Coder Next is a better alternative?
It’s a mixed bag TBH. The issue is that Qwen 3.8 27B is so good.
It depends on the quant you are running though, and the way you are running it. If you have 16GB or less VRAM and a generous amount of CPU RAM, and you are running a regular GGUF instead of something exotic, Qwen next is probably better, with the right hybrid inference engine.
I mean both, for different purposes. They are definitely very different.