alt Hacker NewsGLM-5.3: Frontier coding with emergent cyber capabilities
Comments
This is absolutely still shy of Sol and Fable, but only just by a hair. Ridiculous results. There's still not a compelling economic reason to drop OpenAI courtesy of the ludicrous reset addiction that's taken place, but it feels like we're on the precipice.
How are you all toying with running this kind of thing in a mega quantized way locally? Two weeks out from released weights, but this is still just GLM 5.2 with post-training magic.
I might be just reading my positive bias into that text, but is it possible that it is written less like SV marketing hype trash and more like researchers wrote it?
It does feel like it respects both me and my time.
Thank you, Z.AI. Amazing what difference it makes when the top of your org are actual university professors.
> Mythos 5 remains well ahead at 181 and 247 tasks. The pattern across the three is consistent: the further up the exploitation chain a benchmark sits, the wider the remaining gap to the closed frontier. Capability is growing fastest exactly where we are furthest behind.
I appreciate they don't just take the opportunity to self-glaze.
Same image->html test as I showed in the Gemini 3.7 flash thread. Note that GLM isn't multimodal, but it still was able to generate something similar-ish by writing a python script to inspect the image and extract elements from it.
Original images: https://image.non.io/neonRamenDesigns.webp
GLM 5.3 build: https://html.non.io/neonRamenGLM5.3
Opus 5 build for comparison: https://html.non.io/neonRamen
For having no vision, it did a tremendous job. I'm pretty impressed it was able to extract so much detail.
The Opus one is still significantly better, but that's to be expected since it's multimodal. Curious to see where a future version from Z.ai lands on this.
> Scaling post-training is all we did for GLM-5.3.
Love this opening line. And wow, great results.
> As agent capability improves, much of the difficulty in scaling post-training moves from the model to the environment.
One htought I had; if The chinese allow unfettered access to cyber capabilties while th US does it's best to neuter it's model releases, from China's point of view they have the US all tied up in knots dealing with problems they don't give people the tools to solve. China giggles as it watches the US under threat from people using it's models. The US is restricting citizens from owning this particular kind of weapon, while China is handing it out to the wrolds citizens freely. It feels like the US would only come out worse overall?
Apple will release M7 MacBook Pros / Mac Minis next year, and they will be able to run free LLMs locally at native speed. All software developer notebooks will be replaced to run local models, saving a lot by cancelling Claude Code subscriptions. Developers win. Apple stocks will be rocketing. Everything else will go down. You're welcome.
OpenAI and Anthropic need to just go ahead and give people access to the cyber models.
Otherwise we have a world of attackers using open and closed source models against a much smaller group of maintainers that are likely heavily dependent on Anthropic and OpenAI and for whom it may not be a simple matter to just get approval to start using the open model flavor of the month.
This will be roughly on pair with Kimi K3, but using a third of its parameters.
Just 4 weeks ago the "Kimi K3 moment" was seen as a threat to Closed AI and in less than a month Z.ai have cut the parameter/RAM barrier to a third.
Congratulation to Z.ai and all the hard working Chinese researchers who are quitely boiling the frog.
Their coding plan switched to credits, didn’t it? What are the rate limits like, compared to Anthropic or Kimi K3?
I remember trying their Coding Plan out before the change and the 5 hour limits felt too restrictive then even for light/medium work, especially cause of the whole peak and off-peak thing: https://blog.kronis.dev/blog/z-ai-s-glm-5-2-is-a-great-model...
Nowadays, I’d probably go with their Max plan if the rate limits are okay? Anyone using them now?
Oh also unrelated but ZCode was surprisingly good, which is surprising for a tool that came out of nowhere - even some of the critiques in my blog post have been patched out. Sadly they don’t support using Claude Code as an agent so can’t use it like Paseo or Kepler or Agent Orchestrator.
Missing multimodal again?
It is so valuable in practise to be able to have the models see screenshots - I guess if they aren't in the benchmarks then nobody will focus on it. But it completely nixes these for some of my main use cases.
Meanwhile, my OpenAI TAC application lingers in a total limbo. I suppose I'll switch to this at some point.
" a judge agent then attempts each task to verify that it is actually solvable "
I understand you need to verify the goal is achievable. But if the judge agent has the same goal as the training agent (solve), and both are of the same model, then aren't the judge and the training agent doing the exact same thing? What is the point then? Can someone explain this to me.
People familiar with the topic, how will models continue to get better? Post training it seems? Labs have already used up internet-scale data, so are there any limits to architecture improvements and post training or can we expect this trend to continue? ByteDance is training a 10T-parameter model. Here, GLM 5.3 outperforms models 3-4x its size of roughly 700B, so parameter count doesn’t seem to be a direct correlation anymore.
Such an interesting times we are in,
We just had amazing releases this past two months
kimi k3, glm5.3 qwen3.8 and now glm5.3
These open models are getting really good
I'm not that up to date with the latest AI developments, but I noticed that this article seems to use "Cyber Capabilities" as a shorthand for the model's ability at cybersecurity tasks? Is that now an established expression, same as "crypto" now refers to cryptocurrencies rather that cryptography? Because "cybernetics" actually means something different (yeah, old man yelling at clouds, I know)...
Google: Here is the next iteration of our flash model series, with a discount. please use. thx.
Z.ai: Here is our next iteration, neck and neck with Fable/Sol. weights releasing in two weeks.
A flood of releases today, really difficult to make out for someone who does not use or test all these models on complex real world use cases as to how people decide which ones to use (besides price)
Love these open source models keeping close source models honest.
Wohoo. Congrats to team. Been using 5.2 for a while for hobby use and it's been solid - smart enough for my needs & I'm on a grandfathered plan.
Nice to see a commit to open weights straight off the bat
Musk: Open Chinese models will rival Fable 5 in Q1 2027
JieTang (Founder of Z.ai): It won't take that long
Unless its multi-modal and can deal with screenshots - its not really usable for a lot of coding use-cases.
Look, GLM, Kimi, Deepseek and Qwen should just join forces and come up with THE model that will beat the frontier lab models even just for the benchmaxxing perspective - all just to create hype and chaos to derail the trillion IPO conversations surrounding OpenAI and Anthropic.
This looks very interesting indeed!
So, "cyber capabilities", whoa there horsey, what the fuck is that? Are we making up words or are you trying to court the black hat crowd?
No Hugging Face link yet. I wish they would release it under a true FOSS license.
Kimi and QWEN are now moving on to a restricted-usage license, which, although is still better than the proprietary American models, is a step back from the open source Chinese LLM culture.
The capabilities of open models approaching or meeting that of SOTAs is good in every way except for our short-sighted economic reliance on their success (in the US at least).
Feels like Fable's edge ended up just being long horizon task scaling, which post-training seems to achieve as seen here. Wonder what the next frontier is? Improvement in specialised tasks or computer use?
dario must be writing another angry essay arguing why his closed model AI is too dangerous to be used by others.
I was extremely impressed by GLM 5.2, although you could definitely _feel_ it was a bit behind Opus 4.8 at the time. Eager to see where GLM 5.3 is at.
> The model weights of GLM-5.3 will be publicly available soon in two weeks.
Incredible numbers, will have to wait and see how it actually performs. The timing of GLM updates are always suprising
They're taking security seriously with this one, with their own disclosure page, like Anthropic did for Mythos. https://cvd.z.ai/
Can't wait to see it in practice.
The model releases just don't stop!
> Open Source: We will release the weights in two weeks after launch, once safety evaluation and hardening are complete.
What safety evaluation? What safety hardening? They already evaluated it and found it to be highly capable at exploiting security vulnerabilities. So we know it is not "safe", and they don't seem to plan to do anything against it. What could be more dangerous than hacking? Biological weapons research? I don't think Chinese labs are doing anything against this either.
Their own hardness (ZCode) seems to be a GUI, which doesn't work for me. They say they support other harnesses. However, it seems like I can inject the plan into other harnesses, like Claude Code[0]. Does anyone who's been using GLM models for a while have a strong feeling for if it does better in some harnesses than others, or should I just use my favourite harness?
just imagine the world without these open weight models - we'd probably have to reverse mortgage our homes to pay for tokens to those trillion $ companies to have access to their models.
[total rewrite: their subscription code is buggy. It takes a while for paid subscriptions to show up, and for upgrades to take effect. Original comment was whining about this]
An I the only one who was disappointed with GLM 5.2 after all the hype? It was thinking forever and sometime just stopped mid task.
[dead]
[dead]
[dead]
[dead]
Apparently they are scanning OSS and popular software at scale and disclosing the vulnerabilities they found: https://cvd.z.ai/
Most of these are under embargo, but it seems there are a lot of CVE here from a wide range of popular software, many considered critical or high.
I understand the argument of "people are not actively looking", but isn't the cost for such a scan getting lower by the week, and Anthropic's Project Glasswing is supposed to find them quite a while ago?