[UncleOxidant]

> It is almost guaranteed that a 60-90B model can outperform current SOTA in coding tasks within 2-3 years

Given how well Qwen3.6-27B performs for such a small model I think you could be right. I suspect that Google,OpenAI,Anthropic must be looking at the Qwen3.6 models (as well as Deepseek V4-flash, MiMo-V2.5) and wondering if they could make some smaller models that are specifically trained for certain activities - like coding. Smaller, more targeted models would take up a lot less resources.

[svachalek]

The problem is that once you reach a certain level in coding (not particularly high imo, although some would differ) the most significant improvement in your output comes from understanding requirements better and finding ways to meet requirements in productively lazy ways, bypassing busywork that seems necessary but isn't. And that's the kind of stuff you will only find from a generally intelligent model, not a code monkey that's optimized for turning requirement sheets into source code.

[tracker1]

Personally and mentioned in other threads, I feel that we'll see a breakup of domain/context specific models as well as the goliath models in use. The tooling and a classification model will draw out the context and tooling will pass the work between context specific models in order to improve the cost characteristics of the work itself.