Zeta edit predictions are not up to standard

[Benjamin-van-Heerden]

Been using Zeta for edit predictions for a while now. Really rooting for it.

Unfortunately, it is just not up to standard for me. Simple things like renaming a variable in cases where using LSP rename does not entirely make sense, Zeta does not pick up what I am trying to do. Between file context sharing is also not fantastic. Idk, I think you need to step up from Qwen 7B, it is not powerful enough.

I am not suggesting a something ridiculous 100B+ model, but something a bit more powerful e.g. Qwen 32B. Maybe a larger context window model is also something that needs to be considered, idk, certainly a more powerful model is, I think, not negotiable.

Had a look at the dataset you guys are using. Many languages not even in there. I would be happy to set aside some of my time to help in a crowdsourcing capacity. E.g. Zed has a bit of a cult following in the elixir community since Jose is using it, but not one elixir example in fine tuning dataset.

To be clear, my gripe is not with edit prediction in elixir (I am fully aware that I should not expect the same performance in a language that is very popular in the grand scheme), I am talking basic Python.

Imo, it is just not good enough to do the "basic" things I now expect from an edit prediction provider. I think it can be great, but it is not there at the moment.

I really hope I am not being perceived as a dick here. I really love Zed. It has become my daily driver. I can see greatness in it and the team working on it are obviously super talented.
Just think this is something amiss.

[LaneSun]

I think the 7B model size is a sensible choice—larger sizes aren’t very practical. Just imagine switching to a 32B model and then paying for every 1,000 code predictions (rather than valid ones). Wouldn’t that make you anxious, to the point of coding overly cautiously?

Of course, I agree that Zeta should improve—for example, by optimizing context design, datasets, training methods, etc. The potential of a 7B model is far from fully tapped.

That said, personally, when the results aren’t good enough, I don’t always want to think, "Alright, let’s just throw more compute at it," but rather, "Look at the current conditions—there’s still so much potential to unlock." In this regard, I really admire the Zed team.

[fungilation]

don't need or want much higher than 7b for inferencing speed and cost reasons. But it is true that Zeta is based off Qwen 2.5 and that's outdated by Qwen3 which is far more capable

https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct or https://huggingface.co/Qwen/Qwen3-30B-A3B as MoE with only 3b active params should be great ones to move to. With 3b active it'd be even faster