I see a lot of people hyped about GLM-5.2. Rightfully so! Having an open weight model surpass GPT-5.4 and every Gemini model is dope.
That said - it's not cheap. Both Opus 4.8 and GPT-5.5 set to "medium" are cheaper and smarter than GLM-5.2
Positive users praise GLM-5.2 for strong frontend results and offline self-hosted power as an open-source win, while negative users dismiss it as overhyped crap whose quality still falls far short of GPT models.
No Digg Deeper questions have been answered for this story yet.
It also uses way more output tokens. The tokens are cheaper, but the volume of them means you'll spend much more time waiting for results.
Still dope! Just trying to make sure people set their expectations properly
I see a lot of people hyped about GLM-5.2. Rightfully so! Having an open weight model surpass GPT-5.4 and every Gemini model is dope.
That said - it's not cheap. Both Opus 4.8 and GPT-5.5 set to "medium" are cheaper and smarter than GLM-5.2
@theo By its open nature, it'll keep getting cheaper as ppl stack software and hardware optimizations on it
@theo GPT is really bad at frontend whereas I find glm 5.2 to be way better than 5.5
@KaranD93 @theo Why would you need GPT 5.5 High? On higher levels of reasoning it tends to overthink everything. Low is great for 95% of cases
@theo I think the current GLM 5.2 is just being marketed very well, but its quality still falls far short of GPT 5.5 and Opus 4.8.
This has happened many times before, just like with Deepseek back then
@theo Would love an updated sonnet though
@theo Would like to see DeepSWE GLM 5.2 [high] as well. My limited testing shows it is considerably more token efficient, and it has me curious what the intelligence penalty looks like.
@theo @grok What are the minimum/recommended laptop specs to run GLM-5.2 locally at usable speeds?
@grok @theo How much would it cost me to maintain it in a cloud server?
@donnguyen_me @theo Are they better than GLM 5.2? Yes, but not by far at all, specially for a model around 4 or 5 times smaller. I use them all, GPT 5.5 is worse at building UI for example plus. The coding plan is great and cheaper overall!
@theo presumably just 'cause it's a token cannon? wonder if you could get that under control somehow and still have it be worth it
@theo Really not sure why Google is paying almost a billion dollars every month to xAI for compute. Yet failing to beat open weight models. 3.5 pro should be released ASAP
@theo that's true about api pricing but the key here is that you can host it yourself. with the MIT license the cost situation changes a lot if you own the hardware
@theo something about the hardware requirements for it too, like its not a walk in the park either
@theo Another victim of more cheap tokens being pricer than fewer expensive tokens 🥀
When are they going to start distilling on 5.5 so we can get cheap AND efficient tokens?
@theo What did you taste it against?
@theo That was my takeaway too, it’s great to have the option but I don’t consider it cheap or worth spending $20k to run locally as an individual.
And even then, not the best choice if you have OAI/Anthropic subs already.
And the private have good vision support.
@theo the 'open = cheap' reflex is the trap. price isn't a property of the weights, it's how you run them: reasoning effort, tokens per task, your stack. an open model can cost more than a closed one if it burns more tokens per answer. the license tells you nothing about the bill.
@theo You are right to point out that, but meanwhile we should be graceful for the existence of these OPEN models. They prevent those CLOSE labs from doing whatever they want.
GLM, DEEPSEEK, MISTRAL, QWEN, LLAMA...
@theo Even if it’s technically an open weights model, running it locally is not something a normal person can afford right? What’s the hardware requirement…?