[mark_l_watson]

I am mostly retired but I am thinking of restarting a solo products mini-company next year. I have been looking at much less expensive options like Alibaba Cloud, GLM, Kimi K2, etc. There is a recent Stanford study showing most US startups are using less expensive Chinese models, but I think usually hosted in the US.

For now I am happy enough with Gemini and GPT-5 because my usage is so lite that anything is cheap. For many engineering use cases, Gemini-2.5-flash-lite works well enough.

How do you use GLM? With codex —oss? Or, just ‘raw’ with no agent-wrapping coding environment?

[syntaxing]

I use it directly with Claude code [1]. Honestly, it just makes sense IMO to host your own model when you have your own company. You can try something like openrouter for now and then setup your own hardware. Since most of these models are MoE, you dont have to load everything in VRAM. A mixture of a 5090 + EPYC CPU + 256GB of DDR5 RAM can go a very long way. You can unload most of the expert layers onto CPU and leave the rest on GPU. As usual Unsloth has a great page about it [2]

[1] https://docs.z.ai/scenario-example/develop-tools/claude [2] https://docs.unsloth.ai/models/glm-4.6-how-to-run-locally

[mitjam]

Hope you‘ll share your story if you start. Love your book on langchain from iirc 2y ago, it got me going.

[mark_l_watson]

Thank you!

[mistrial9]

> There is a recent Stanford study showing most US startups are using less expensive Chinese models

link ?

[mitjam]

Idk if this is the reference but it’s in the same direction:

„ These days, when entrepreneurs pitch at Andreessen Horowitz (a16z), a major Silicon Valley venture-capital firm, there’s a high chance their startups are running on Chinese models. “I’d say there’s an 80% chance they’re using a Chinese open-source model,” notes Martin Casado, a partner at a16z.“ —- https://ixbroker.com/blog/china-is-quietly-overtaking-americ...