Qwen3.8-Flash-Next is better still if you can run fast enough. If you have a dual R9700 setup you certainly can. That model is even better.

Qwen4-27B has been announced but not released yet. I'm super pumped for it because I already use 3.8 as my daily driver at home for all my personal stuff, so I'm definitely happy to take an increase in capability.

There is clearly still room for improvement in local models on consumer hardware. With the Qwen 27B models, If you have at least a 5070 Ti I think you can get away with running a small Q4 quant if you use KV cache streaming. The 24GB cards can run Q4 comfortably. If you have a 32B card you can run Q6 comfortably. If you have 48GB ~ 64GB of VRAM you can Q8 comfortably. Using llama.cpp Vulkan let's you pool VRAM across cards (even AMD and NVIDIA etc), so my machine has a 5060 Ti and an R9700.

A dual R9700 rig is really the sweet spot right now with the vLLM-radiance fork. If you can swing a 5070 Ti in there as well to retain some CUDA access, then all the better. That's basically the equivalent to spending 2 years on a subscription, but gets you a system that can run Qwen-3.8-Flash-Next and of course the even more capable Qwen4-Flash when it releases. At the end of the two years it'll run even better models I'm sure.

I'm all in on local now.

The optimal move is to retain the minimal access to SOTA models on the $20 plan, and for anything your local model fails at, use SOTA as the backup for either planning or debugging.

This way you're not actually at any disadvantage in terms of capability. You also don't need an advantage, you need to complete the tasks you care about. Eyes on the prize.

RTX 3090 came out a long time ago and it may be 'outdated' at this point but still banging like a champ for anyone who bought one and becoming increasingly more capable as new models unlock it's potential. Hardware hasn't changed much, but what it can do certainly has.

That they are expensive and climbing doesn't negate my point if the cost of the subscription over how long you plan to keep it is equally or more expensive than the GPUs. You can put together dual 5060 Ti or 5070 Ti systems to run local LLMs too. You don't need to splurge on a 5090. That's a bad option at this point.

I don't really buy it actually. There isn't really anything meaningful you can learn with how to use LLMs/agents that has a half-life greater than a few months at this point, so you can just start doing it at any point in the future and not be meaningfully left behind. On the other hand years of letting your actual engineering skills atrophy will have a negative effect on you. I've been witnessing the effects of this. Going back to more coding by hand with AI-assistance circa the 2023 era as a happy medium. I think this is the sweet spot. Full agentic engineering has nasty failure modes and in the long-term is kind of a bad option for basically everyone. I say this after having done it for almost a year at this point, and transitioning away from it now.

Ah, so you're still a few months out from the "yeah, maybe I don't really love this and maybe it won't ever actually work as well as I thought" turning point.

You wouldn't download a car.

We need to put more money into human alignment research.