Given that DeepSwe is one of the very few coding benchmarks worth taking a look at, this achieves rather excellent result at it (not far from opus 4.8).
From looking at the results and my own impression of 5.1 and other models, I think this is the best Chinese coding model by some non-insignificant margin.
Yeah, that's what I mean anyway. Each model has certain design tropes it repeats everywhere, and some of them are very old-school or not really UI best practice.
And then the more ambitious cases where you ask for a feature without being prescriptive with UI needs, the end result is sometimes atrocious with weird font use, colours, etc.
Zhipu AI is founded by a superstar Tsinghua professor, did an IPO in January (Hong Kong stock exchange) hired half it's past research lab and it's stock is >10x since. This is not a "just distill Claude" thing.
HuggingFace says this model has 753B parameters, which will need a lot more RAM than a maxed-out MacBook Pro. With 40B active parameters, running from SSD would need patience.
I’ve wondered for a while if anyone is working on very wide channel parallel (kind of like RAID 0) SSD for this purpose. Couple that with a tensor processor and that would be interesting.
Given that DeepSwe is one of the very few coding benchmarks worth taking a look at, this achieves rather excellent result at it (not far from opus 4.8).
From looking at the results and my own impression of 5.1 and other models, I think this is the best Chinese coding model by some non-insignificant margin.
I've been very pleased with it's performance over the last few days.
It's definitely not near Opus 4.8 level but it's very impressive nonetheless and it does do design extremely well.
> it does do design extremely well
Better than Opus?
I don't know what people mean when they say design lol, is it for frontends?
Yeah, that's what I mean anyway. Each model has certain design tropes it repeats everywhere, and some of them are very old-school or not really UI best practice.
And then the more ambitious cases where you ask for a feature without being prescriptive with UI needs, the end result is sometimes atrocious with weird font use, colours, etc.
Zhipu AI is founded by a superstar Tsinghua professor, did an IPO in January (Hong Kong stock exchange) hired half it's past research lab and it's stock is >10x since. This is not a "just distill Claude" thing.
IPO within year of founding?
It it normal for startup in China?
Blog post with more information including benchmarks: https://z.ai/blog/glm-5.2
Discussion (765 points, 3 days ago, 489 comments) https://news.ycombinator.com/item?id=48518684
Looks like their code subscription is sold out. Clicking on the "Subscribe" button just does nothing. Pay-as-you-go API works well.
If I have a fully maxed out MacBook Pro, would it make sense to just switch from Opus 4.8 to this? I've never tried running local models for coding...
HuggingFace says this model has 753B parameters, which will need a lot more RAM than a maxed-out MacBook Pro. With 40B active parameters, running from SSD would need patience.
For an fp4 quantization it should fit with room to spare for KVCache
Aren't Macbooks limited to 128GB RAM?
FP4 would require >350GB RAM + KV cache, so no.
Oops, you’re right. My brain understood Mac Studio for some reason.
I’ve wondered for a while if anyone is working on very wide channel parallel (kind of like RAID 0) SSD for this purpose. Couple that with a tensor processor and that would be interesting.
There's talks about HBF. F for Flash -- HBM packaging and bus width, but using NAND memory.
e.g. https://www.sandisk.com/company/newsroom/press-releases/2026...
I have already tested this and it is impressive.
non-x blog post: https://z.ai/blog/glm-5.2
And previously: https://news.ycombinator.com/item?id=48518684