79 comments

[ 1.5 ms ] story [ 71.5 ms ] thread
Is it a coincidence that both MiniMax and Z.ai are releasing frontier open weights models right as the USG is trying to impose a cap on model capability offered to the public?
Released at the exact same time, 5:21 pm (Chinese time), as when Anthropic received the letter from the government banning Fable, and explicitly citing other models becoming unusable.
(comment deleted)
With deluge of Chinese models popping up recently, I believe there's a few issues one needs to evaluate before deciding to use these models:

- Ethics. As known, ou American frontier AI companies are incredibly ethical. And I have yet to see any interviews or blog posts by Chinese companies where they talk about how they are ethical, or at least credible HN comments about it.

- Safety. Do they covertly sabotage or at least refuse to answer questions that could help cyber- and bioterrorists in their nefarious purposes? What about ML-related questions that could help terrorists create AI models without guardrails?

- Child safety. This is especially important with "free for all" open-weight models, most of which are Chinese (ever think about why that's the case?). How are we going to do age verification and KYC with models that anyone can just download on their computer?

- Intellectual property theft. How can we be sure that no output of our American frontier AI models was used while training these Chinese models?

Frankly, there's a plethora of other issues I don't have time to get into right now. Personally, I believe distribution of Chinese models in the US should be paused until they are required to submit models to the government for review and evaluation, to make sure they are made to Anthropic/OpenAI standards.

We need legal grounds for that.

Write to your congressman, congresswoman or congressperson and urge them to stop proliferation of dangerous non-American intelligence. This is a matter of national security and needs to be acted upon as soon as possible, preferably before IPO.

The funniest thing about this post is not the fact that some people took it as anything but satire, but that it’s likely very close to what the true believers at Antrophic actually think.

Ah, those wacky terrorists and their non-aligned models, trained on copyrighted data to boot. Remember, the only thing that stops a guy with an evil god-in-a-box is a guy with a benevolent god-in-a-box, and only Antrophic can lead us to the second one – but only if we act together as a nation and ban those subversive open weights models!

I used to buy only American Ethics, but Chinese Ethics are becoming pretty good lately for the fraction of the price.
(comment deleted)
The solution is tarrifs. Require 3 american tokens for every imported chinese token
This is brilliant, but you should have added some <joke> tag or something. You'll be confusing a lot of people and I really can't blame them. I think I've already seen all of these arguments used here seriously in one way or another.
Seems like there's no official blog post with benchmark results yet. But I'm once again thankful for the Chinese AI labs for being open with their work and contributing it to the world under permissive licenses like this. The Fable 5 fiasco is just another reminder of how valuable these things are to have.
Based on my first impressions it's about 6 months behind the frontier labs. So very similar to Opus in January.

That is, pretty damn impressive and very useable. When it comes to architecture or complex problems it does noticeable worse but I don't think anyone expected anything else.

One particular interesting strong point seems to be design and user interfaces. It does seem to punch above it's weight there but that might just be personal preference.

Releasing a model without benchmarks seems to say the model is probably bad...
(comment deleted)
I just ran a report from a project I'm working on that uses a mix of models, and GLM 5.1 trumped Sonnet over the last week, so I'm excited to now turn on 5.2. This is based on completion only - not quality, but that includes passing a huge test suite, and Sonnets failure rate was surprisingly bad...

What I've seen from 5.1 for things like planning has certainly not read as impressive as Opus, and often even as Sonnet, but it's been a strong and steady work-horse that's just kept on actually delivering progress.

It's also a reminder that as soon as Chinese models take the lead, they will switch to closed source too... so let's not be complacent, we need stronger, completely open data models, open source code, etc. to mitigate this risk
I wish they would write a blog post about capabilities of this new model, what to expect from this model, is it cheaper, is it faster or does it have better quality in the outputs.

But still, thank you for the release

maybe wait til monday guys
Weird, z.ai does not resolve for me. Is there anything special about that domain?

https://z.ai

If you have systemd-resolved, it tries to validate DNSSEC by default and replies with SERVFAIL if it fails. Same happens here, I go through some privacy focused DNS servers and they sometimes remove the signature.

$ resolvectl query z.ai

z.ai: resolve call failed: DNSSEC validation failed: no-signature

This release was rushed to hang on the coattails of the Mythos drama (“hey, sorry you can’t use Fable, but try us while you wait this weekend!”) I think they planned to release next week, hence benchmarks not all being ready yet.
Could be, but AFAIK it was similar with other glm releases. Just a Twitter post with blog post coming later.
It would be so extremely awesome if this ai would have been a Claude killer alternative and 90% of Europe cancels Claude subscriptions and subscribe on this one. It would be the dumbest move of the year by the US.
For personal use I already did a few months back. Dario is more competent than Sam, but even shadier (IMHO).

Anyway, switched to Openrouter through forgecode (or pi/opencode, the jury is still out on this one).

It will take a while, but I believe that also businesses will at least hedge against US companies basically being forced to geo-fence their models. For now is Fable, but they can include any model at any time.

I'm actually interested in doing that.

What would be the most favorable model/company to move to for scientific programming and engineering questions?

Given the US government’s latest stunt with Fable, this is looking more and more like the future.

Can’t rely on strategic products if they’re gated by capricious actors.

Open weight models are basically immune to that

> Open weight models are basically immune to that

Somewhat. The US Gov can make it illegal to transact with, download, use, etc. foreign open weight models.

Of course, enforcement will be difficult for individuals (businesses will comply by default, and they would all be pulled off Github and other US based hosting locations if they went the sanctions route). But, we are also quickly going down the road of frightening levels of mass surveillance, which could aid enforcement.

The Fable situation sets a very dangerous precedent, and I'm not looking forward the future here. We are losing the fight for information and computing freedom.

It’s very likely the Chinese go dark too the second they have parity / lead
You criticize the government, perhaps rightfully, but give Anthropic a pass. They are the ones fueling this bullshit. Downgrading your results without telling you. Refusing your requests in the name of “safety”. Even if the government didn’t make them pull the model for foreigners, we’d still be in a really shitty situation because Anthropic is really shitty.
will simon do the pelican thing for this as well
Announcement from the founder of Z.ai:

“ GLM-5.2 is Fully Open, Frontier Intelligence Belongs to Everyone

Today, the sudden restriction of certain frontier models is deeply regrettable. At a time when access to frontier models is abruptly cut off for non-technical reasons, we are even more convinced of one thing: science should be global.

The path to AGI (Artificial General Intelligence) must never be enclosed by high walls. We have always believed that AGI should be the cornerstone for all of humanity to collaboratively explore the boundaries of intelligence and solve complex challenges, rather than a privilege monopolized by a few rules and subject to revocation at any moment. In the face of external blockades and restrictions, our attitude is one of radical openness. Frontier intelligence must remain open-source, accessible, and buildable, serving every dedicated developer.

GLM-5.2 is Zhipu's most capable open-source model to date. It not only supports a truly usable 1M context window but also maintains a continuous lead in the independent completion of long-horizon tasks, providing solid foundational support for building complex agent applications. It also continues to be our main engine for creating the strongest domestic coding model.

Tonight at 5:21—at this special moment—GLM-5.2 will officially be available to all GLM Coding Plan users (including Lite / Pro / Max). The API will also go live next week.

A step closer to frontier intelligence for everyone. The future of AI is open, and it is for the people. ModelKey: GLM-5.2”

https://x.com/jietang/status/2065784751345287314

AI seems to be renewing and amplifying our cultish behavior as a species. AI is not going to save us from ourselves.
Is there any indication of what compute resources this will actually require (in its various incarnations)? Does it incorporate any of the optimisations pioneered by Google (such as TurboQuant, MTP) or some other original innovations to make the frontier quality realistically available to local users?
[stub for offtopicness]
Curious what people's experience is with these models. Anecdotally I tried these out earlier in the year and found it struggled with pretty basic full-stack coding I was doing, when Sonnet 4.6 and Haiku 4.5 didn't break a sweat. Was hoping to use it while my Claude usage was resetting but was disappointed.
It's great that we are getting so many open source model releases, but I just feel like SOTA models will always be in the hands of the big players. The hardware requirement to achieve SOTA are just too steep.

My alternate universe would involve some sort of decentralized investing scheme to build data centers running massive open source models that could compete on some level with Anthropic, OpenAI, etc.

I don’t think this stands for General Linear Model.
In the last few days, Chinese labs have given us MiniMaxM3, KimiK2.7 and now GLM5.2. Meanwhile US is censoring models. Reads like fiction.
Crossing fingers for a 5.2 flash release - it’s been a while but I still feel like 4.7 flash is one of the strongest local coding models
I wish the torrent would come before the announcement. Doing it the other way is playing with fire.
I used to use GLM before I knew about coding subscriptions and it was okay. I've tried every version since 4.6 and this one is doing a great job a spec-implementation runner. If I had to guess... somewhere between Sonnet and Opus in terms of quality. Z.ai's issue has been service reliability. So far so good on day one.