Context - Deepseekv4 is freely available to download you can host your own and sell it keeping the proceeds and it rivals Claude Opus 4.7.
"Thank you for your attention to this matter"
It's worth noting I suspect a key reason Chinese companies are doing this is, in part, tacit encouragement and logistical enablement from the Chinese government. Playing spoiler by nerfing the valuations of over-inflated U.S. AI leaders is a decent strategy given the current GPU disparity.
Having the software stack get dirt-cheap and multi-vendor is good for hardware players. I'd argue there's even huge value in developing tooling that's less nVidia-centric for the health of the overall market, and the fact China is filled with other hardware manufacturers that want a bite at that margin and insatiable demand.
good luck getting a machine that can run its specs though. Even flash is goign to require ponying up 5-10 grand to run the minimal specs for it. The vast majority of people will find their machine falls behind as tech progresses long before they get a return on that investment. That said, it does mean there will be a healthy market for "generic providers" in the AI landscape with these open weight models.
Investment is not basic math. Its also dependencies to US companies, trust etc.
We’ve all seen how the “math” on so much of the AI business sector literally doesn’t check out, and here there are: still ballooning, still making deals, still directly crafting laws through political influence, still taking over damn near every user space.
Politics at a high enough level lets you play a different game with different rules.
Politics at a low enough level lets you do the same actually, but we usually call that civil unrest, guerrilla warfare, or collective action depending on how many of which group is defying which “rules”.
It’s very easy to get used to the guardrails and guidelines around us when they persist and succeed for decades, but they are much more fragile than they appear.
That's any machine that can physically host the weights and context. You'd need a highly-specced machine for better performance and throughput, but it's not a requirement as far as literally executing the model and getting output.
If you know exactly what you want and when you need the results, calculating a hardware floor becomes deterministic. If you don't need the results right away, or if you're comfortable gluing together results yourself, pretty much any box released in the past decade can be doing work for you.
For a quick experiment grab a few images and a one-line prompt e.g. "describe these pictures", and bounce that set off every model/quant you can reach. If you record quality, rate, and cost of each request there might be regions of
`"good enough", "good enough", "electricity + sweat"`
in the resulting spreadsheet. And this is multimodal. Single-mode classification is dirt cheap.