I'm pretty happy with Kimi 2.7, and I've heard good things about GLM-5.2 too. Honestly, I haven't even exhausted the ways I can use these two models yet. I could easily keep getting value from them for at least another two years, creating new coding projects and tinkering with what they can do using skills and tools.
At this point, I would happily pay €250 or €350 for a PCIe card capable of running a local model with similar capabilities, something like what the guys from Taalas (the creators of chatjimmy.ai) are working on. I don't even need crazy speed, just affordable, energy-efficient hardware would be enough.
I grant you that my use case (affordable and easy-to-use on-premise inference) is a bit niche. But I also see a massive opportunity here for smaller, tool-capable models in things like Home Assistant and smart home appliances. I know this completely goes against current market trends of "Everything as a Service" and subscription-based models, but the home appliance market has always targeted users who run away from those kinds of ecosystems. In the end, I just don't want to pay a €20/month subscription to do the laundry with my own washing machine.
We mentioned Taalas, but they are not the only ones trying to strip away expensive, power-hungry VRAM in favor of more rigid but purpose-built, energy-efficient solutions. Check out this article to find more companies innovating in this space. Imagine a small PCIe accelerator, similar to what Coral AI or Hailo did for computer vision, but completely optimized for running local LLMs on our own terms. Wouldn’t that be cool?