Sort:  

Qwen 3.5 27B is probably the best choice for local, but it will be slow but it's a solid model with much lower memory demands than most.

and 1M context window

27B is only 262K context window.

80b has 1M :)

Are you talking about Qwen3-Coder-Next 80B? It's only 262K too, but either way these start to get context rot after 100K.

you can enable it to 1M ( but is then slow as fuck). Well and it makes things up. Thats why i think memory files could be the thing to solve it better on infrastructure. But i am still in fuck around find out on that.

btw since you are also deeper into it, i experminted with memory files ( like claude) a bit. Any experience with it? So far with stuff like opencode + plugnins like rp1 or others it works well to be always up to date.