Qwen3.5 is able to output click coordinates and bounding boxes just fine, as values normalized to 0..1000, I’d hope Qwen3.6 didn’t loose this capability.

I've had lots of success with generating coordinates and answering questions using the UI-TARS model https://github.com/bytedance/UI-TARS.

This sounds a lot like another hacker news posted in the last few days. The same problem image generators have with a prompt like, produce numbers 1-50 in a spiral pattern and it can't count properly. But if you break it into a raster/vector where you have it first produce the visual content and then a SVG overlay, it's completely capable.

Have you tried doing a two step: review the image, then render a vector?

>but it didn't perform well in our coding and reasoning testing

>Comprehensive evaluation results at https://gertlabs.com/rankings

But if you go to the linked site, it seems like the only thing that's part of the evaluation is how well the models play various games? I suppose that counts as "reasoning", but I don't see how coding ability tested?

GLM-5.1 does not support image input.

This may be a strange request, but is it at all possible to include Cursor's Composer models in your tests?

I think the point is to use them both with GLM 5.1 delegating vision tasks to GLM-5V-Turbo

What version of Kimi was that using? Do you have any specific insight on Kimi vs GLM in real world scenarios?

> I've been using GLM pretty much exclusively last 6-8 months. I have access to Anthropic and OpenAI models and others. I always keep returning to GLM, it isn't the best

Very interesting. What sort of tasks do you use it for, and what client do you use?

When you want to use a custom client and a coding plan to control costs (daily use, a few hundred USD/m budget), this is the landscape:

- Anthropic/Google: Deterring custom clients actively

- OpenAI: Grey area.

- Z.ai: Technically only allows clients in their (large) approved list of clients. Likely won't actively ban custom clients.

- Moonshot: Seem to allow custom clients?

- DeepSeek/Alibaba: No coding plans at this time

This... doesn't make sense. Why would they use a quantized model when load is low and the full model when load is high???

I have a subscription and I have not seen any difference in performance during on/off hours. What exactly are you basing this on?

Do you have proof for this?

I hear a lot of people complaining, I am on their Max plan, I never hit limits, use it non-stop and overall it has been fantastic experience.

I was one of the people just absolutely in misery when the GLM-5.1 model dropped. It wasn't quantized, I don't think, but it had some very gnarly issues where it would hit a context size, then seemingly try to quantize, and fall apart. It was unusable. It went from being an excellent model all the way to 200k, to being only 60k before it couldn't write in sentances and definitely couldn't tool call, to being 100k, to 120k. It was terrible, and I was so sad they had made my subscription so much worse, it felt like. https://news.ycombinator.com/item?id=47677853

But very shortly after this submission/release of 5.1, after a mass pouring out of sadnesses, they fixed it. Things have been back to absolutely amazing. I joined right before 4.7, and 4.7 was incredible. 5.0 was fantastic. 5.1 has been a dream. GPT still catches a lot of stuff and is smarter, but man, GLM-5.1 is so capable, and it's frankly often a better writer, often better understands and captures purpose and notion, where-as GPT often feels dry and focused on narrow technicals. I really appreciate GLM-5.1.

And I'm really glad Z.ai fixed the absurd damage they had in their systems. I do suspect they were trying to dynamically quantize as the context window grew, or some such trickery. It was not working at all, but somehow it tooks months to fix.