We built a complete production-grade inference service from scratch on a cluster of more than 100,000 Chinese-made AI accelerators. All production inference for GLM-5.3-Flash runs on this system.
very few people comprehend - how much of an asteroid level event for western AI labs this is.
china has cheap abundant power, now they can make their own inference chips (which was supposed to be a chokepoint), their models yeah can be 6 months behind the frontier - but most people don't need frontier models - small models r more than enough.
my only wish was labs like Mistral would make their own inference chips or partner up eg with established / new chip makers or companies like Oxide.
I was gonna ask how people found their coding plans, and realize, have they massively ramped up the prices? Seems the middle plan is ~$80/month now, didn't that used to be like $20/month? Cheapest plan is ~$20/month currently.
They must have hit really hard scaling limits if the prices were hiked so much so quickly.
It's hard to know, since no one advertises the actual token limits (partially cause they're prolly complex / adaptive). So it seems much more likely that they just offer different pricing tiers than you're used to. Like, the $80 plan is still ~$80 of subscription quota, regardless of what else is offered.
For [API usage](https://openrouter.ai/z-ai/glm-5.3-flash#providers) they charge a bit more than the very cheapest providers of GLM-5.3-Flash, but not so much that a big price difference would make sense.
I paid $360 annual for Max plan and currently averaging about 1BN tokens a day with their frontier GLM-5.3 model. This was clearly unsustainable for them and they've dropped this package.
Well, there's essentially two major ways to use these models: Pair programming or fully autonomous fire-and-forget code generation. The second strategy needs essentially zero input, so the number of tokens you can blow is practically only limited by API speed.
Share the resulting code from any one of those please? I've tried so many times to find a setup that facilitates parallel work + high quality results, but it's just impossible regardless of harness or model. Leave the agents alone for too long, and the entire thing just balloons out of control, and next you know you're sitting there with half a million LOC where 80% isn't even needed.
>I was gonna ask how people found their coding plans
Very good - but I'm on a legacy plan. And coming up on a renewal that would put me on the watered down current plan. But with 50% legacy discount think it may be worthwhile. If I go to a competitor I'd be paying market rate.
>They must have hit really hard scaling limits if the prices were hiked so much so quickly.
Not really scaling - their plans were initially comically subsidized even more so than what the western providers are doing. More advert for an upstart than commercially priced.
Way to restrictive in terms of tokens provided. I am on their largest plan, and quickly run into their limits. And that is using it selectively in addition to codex.
The max plan will provide ~1,100 USD of GLM-5.3 or ~260 USD of GLM-5.3-flash per month for 168 USD. I can personally attest to these numbers through omp (~97% cache hit rate).
Unless you are able to highly parallelize (your work, you won't be able to hit your hourly or weekly quota using the flash model simply because it's so slow.
They give you ~3x more flash tokens, which maybe comes out to ~2x more actual work after accounting for the extra thinking it does to achieve the same result. The mental model, for not getting angry, is 5.3 is fast mode by default, and you can disable fast mode for 2x the work output at 1/3-1/10th the speed.
They're serving me 5.3 at ~40 tok/s and 5.3-flash at 30 tok/s (according to omp).
That table assumes cache hit rate of 95% or better. Am I understanding this correctly that people really are doing such repetitive prompts (compared to each other, across the concurrent user base at that time) that only 5% or less need actually be computed by the intended LLM?
Well, other than the infrastructure they got from illegally routing millions of paying customers' requests through Anthropic's Opus 4.8 in a distillation attack...
Are you joking...? Sorry if so! Just in case: It's illegal in both the PRC and the USA.
In the PRC, they[1] leaked tons of national secrets on the PRC's latest AI campaigns, the inner workings of their "opinion monitoring" (read: performative panopticon) and "stability" (read: violent oppression) departments, Chengdu's whole CCTV network, direct-energy weapons plans, espionage activities in Syria to hunt down Uyghur refugees, and god knows what else that Anthropic didn't divulge to us common folk.
In the US, it's very clearly an attempt to rip off a competitor. I'm not sure how else you could possibly see it. Even if you're a distillation fan in general (which A. why and B. plz don't), they did this through a network of Japanese and Signaporean shell accounts, presumably at least some of which were abusing Anthropic's subscription service in a ToS double-whammy, as it would be exorbitantly expensive otherwise. They also had to hack around Anthropic's API to get CoT traces, which seems impossible to explain away as anything innocent.
I've been beating the "China isn't necessarily an enemy, it's gonna take us all to handle AI" drum for literally years, but this attack was just... gross. Gross in scale and gross in arrogance. Not a good sign for the dawning alignment crisis, to say the least :(
TL;DR: Use these services if you want, but know that you're supporting aggressive escalations and companies that very clearly don't give a flying fuck about violating the law, much less your ToS. So... buyer beware, I guess.
[1]: For clarity, Z.ai was not alone in this, nor were they most egregious attack -- Moonshot.ai (kimi) took that coveted prize. DeepSeek was involved, too.
I'm always wondering when "distillation" comes up how feasible it is, or if it's just BS.
The Antrophic article mentions "16 million" conversations, GLM models are in the 700-300 billion parameter ranges and while the frontier sizes aren't know but Gemini suggests Astra and Mythos are at around 10 trillion. That'd amount to extracting 40k parameters per conversation without a lot of errors if it was just a distillation (from an unknown source/algorithm as opposed to distilling your own model).
Now, I can imagine these conversations being used as a verification step that they're not missing stuff in their training, and that their models are capable of most of the same things, but that's mostly confirming that they've stolen the same data from the public as Antrophic/OpenAI has stolen already.
Or am I missing something here that makes real "distillation" feasible?
That is such a canard, IMO. FWIW, Anthropic and OpenAI encrypt "thinking" token outputs in their models, while Chinese labs don't. If anything, it's more likely that everyone is using open-weight models in their synthetic training data generation pipelines. It's way easier to distill from logits than it is to distill from hard tokens.
No real reason to respect any terms they might want to impose. Besides, if you want to break TOS, just have an agent do it; "everyone" running these things agrees there's no corporate or moral liability for what your AI does.
I'm not defending their actions, but we should be clear about where the law currently stands: Anthropic was found to infringe because of the torrenting, not because of the training.
If you understand what they have achieved here, then the notion that they are bottle-necked on training data is absurd.
I wonder how you imagine that China built their own space station? Reliant on using American made duct tape, perhaps?
Do you realize how reasoning models are being trained nowadays? You design/build simulation environments to run agents in, with the environment providing the RLVR "verification" scoring. So why won't Ziphu use GLM to build their own RL training environments? Do you think they are not doing this?
> Today, GLM-5.3 has become an indispensable daily coding partner for everyone on the team, and it is moving steadily toward replacing us. If this trend continues, given enough compute and enough time, its endpoint is a system that can design and train its own successor entirely autonomously. This is known as Recursive Self-Improvement, or RSI.
You're tragically misinformed; it isn't. Several metrics are actually growing exponentially. But if you want emprical information, you can just have al look at the nature of the late AI incidents.
Ironically, many benchmarks being maxxed out, and quite quickly, so new ones have to be created.
It was not solved. ~OpenAI~ Buckmaster and Alpöge found one (or a few) singularities in the forced version of the Navier-Stokes equations. Then magically 2 weeks later OpenAI found them too. Again, I am not saying this is not a great feat. I am just saying that everyone should be a bit more careful when making statements about RSI.
Buckmaster and Alpöge has found a forced singularity for the 3D incompressible Euler equations (and two other types) building on the work by Diego Córdoba and Luis Martínez-Zoroa with the assistance of Anthropic and OpenAI models. Then OpenAI found a forced finite-time singularity for the Navier-Stokes equations.
As the sibling comment points out, that is incorrect: they solved a smaller, simpler problem, and OAI solved the actual Millennium Problem. But even if they had solved the actual problem and OAI stole it by digging through chats, that hardly supports the "AI has stalled" thesis; Claude played the major role in creating their blowup to a different problem.
How, exactly, does "it was Claude that solved Navier Stokes, not ChatGPT!" get you to "AI has hit a wall and stalled"? That's, not to put too fine a point on it, incoherent, and is just noise thrown into the discussion to avoid grappling with the fact that AI continues to rapidly improve.
> This is known as Recursive Self-Improvement, or RSI.
Some call this "The singularity" (e.g. Hinton).
This is actually a core danger postulated by the, let's call it, "worrying" scenario - see AI 2027 (to be clear, I think its timeline is not realistic).
> Statements dreamed up by the utterly deranged.
Evidently, and tragically, it will take catastrophes to show that deranged are the ones deriding the worried crowd.
This article left me with one immediate question: "WTF is GLM?".
Honestly, I have no idea what z.ai is either (I'm aware of an AI-enabled editor called Zed, but that's under zed.dev), so it's a bit presumptuous from them to assume that everyone is familiar with their product...
Maybe my post sounded harsher than I intended, and yeah, it's probably on me that I'm not familiar with GLM. Actually the other major Chinese LLM Kimi does ring a bell, maybe it's because three-letter acronyms are a dime a dozen and annoy me because I'm confronted with them regularly at work too (people at my company seem to love acronyms), but that's obviously on me too...
Appreciate the clarification. For me it was the "F" in "WTF" that tipped me. Other than that, it's more than fair for you to not know what GLM is. Things are moving so fast that I would be surprised if anyone can keep track of it all. Cheers, have a grand day!
Different angle on the same model: the full GLM-5.3 (744B MoE, 4-bit experts, 434 GB on disk) runs on a single MacBook Pro M5 Max with 128 GB
by streaming the experts from NVMe SSDs instead of keeping them in memory.
One drive gives about 2 tok/s; striped across four drives it reaches 3.5 tok/s with byte-identical output, and our best internal build with a not-yet-published patch does 4.2.
Interesting that the tone of announcements between US and Chinese providers is converging.
GLM has in the past been more technical rather than speculation about future development on RSI etc.
Also curious whether those 100k accelerators are entirely locally made. If that's genuinely end to end on all components including lithography, memory, design etc then that is quite a feat.
Ziphu (who make GLM) use Huawei Ascend processors made by SMIC. Huawei use a combination of domestic memory from CXMT and leftover (pre-sanctions) memory from Samsung.
Just like the rest of the world, including the US (Intel, Micron), SMIC are currently using ASML lithography equipment (DUV, not EUV), but Shanghai Aishengna are now moving into early production with their own DUV machines, with SMIC and CXMT as early customers.
There is also a state sponsored Chinese EUV development underway.
Zhipu has built a production-grade inference service for GLM-5.3-Flash, running on a cluster comprising over 100,000 domestic AI accelerators. The practical application of GLM-5.3-Flash has given rise to a fundamental feedback loop: the model contributes to optimizing the inference infrastructure, while the optimized system continues to serve the model.
Necessity is the mother of invention. The shortsighted protections put on chips, etc., by the US has forced Chinese AI industry to adapt or die. Guess what their response to this fitness function has been? Kudos to Z.ai on their inventions and excellent write-up, which reads like humans wrote it.
Wouldn't it be refreshing if OpenAI and Anthropic were this open, and spelled out how they were using their own models during development and rollout?!
All I can recall reading from OpenAI about what they have actually done in the name of "RSI" is using one of their models to help automate the training process.
US chip export restrictions may actually be an advantage for China's AI Infrastructure. Chinese companies are forced to speed up developing their own AI chips
China themselves recognize this. After Trump relaxed sanctions and allowed NVIDIA H200 sales to China on a case by case basis, the Chinese government stepped in to essentially block it!
In addition to Huawei who make the Ascend series that Ziphu are using, there are also at least a half dozen or so other Chinese companies also making their own AI accelerators.
Look at how China does it. They'll happily sell us everything we want - more than enough of it, cheap enough, to put all of our own manufacturers out of business.
US companies should now be more worried about Chinese companies flooding the market with their, hopefully, very affordable GPU's. The scale at which they can manufacture stuff is unmatched anywhere else. Nvidia can kiss goodbye to their 75%+ profit margins.
Almost everyone knew that these sanctions would backfire within a few years. You can't really put sanctions that have noticeable negative effects on bigger economies. They only work for small to medium economies. I believe sanctions on any economy in top 10 would fail.
pretty sure they'll be fine for a while, between the build out and import bans, I don't expect demand to slow enough to let supply catch up
Nvidia are likely more concerned about AMD taking market share, and I suspect that geopolitics will leave US/China GPUs with largely non overlpping customer bases.
I've got to wonder how long the US will still be able get much of the world to go along with requests like this. Europe is already showing some spine and refused to join Trump's war with Iran, and Italy and Spain have also recently shown Trump the middle finger.
It'd be ironic if Europe ended up banning ASML (a Dutch company) from shipping to US instead!
SCOTUS shutting down Trump's Tantrum Tariffs gave everyone a little breathing room
Mark Carney said earlier today at the EU, "the goal is not self-sufficiency, but collective resilience." It is through collective action that they can find more resilience to our American Antics.
China is gated by not having EUV machine access. They're also bottlenecked by ASML's DUV machine production like everyone else. There are already talks of banning China from even purchasing DUV machines from ASML.
So until China solves the ASML problem, there won't be any flooding.
They've overcome every hurdle to go from nothing to having working machines, so all these questions are presumably a matter of how fast they will improve and ramp up production (initially 5 this year, 20 next year), not whether they will.
China have been really squeezing all they can out of DUV machines, but I'm not sure how much node size really matters for AI competition - more of a cost issue (more chips/power for same FLOPs) than anything, and TSMC & NVIDIA's healthy profit margins need to be considered too.
They already have working machines, and are slated to ship the first 5 early production ones right about now (before the end of 2026), and are projecting 20 for next year.
Small numbers perhaps, but this is happening right now.
What will the numbers be in 5-10 years time? Who knows, but ASML took about 5 years to go from 20/yr to 100+/yr.
Of course politically the world may well be quite different in 5 years time, as may be the AI market.
Same thing with Trump not helping Ukraine and berating NATO. He thought he held all the cards, but now Ukraine has a thriving battle-tested drone industry, UK and France have stepped in to replace the US with advanced missiles and anti-missile systems, stepping up their own production and transferring IP to Ukraine.
Now, the US is left out in the cold with little influence left, themselves now the ones with an anti-missile shortage.
You seem to misunderstand what Protectionism is.
This is not an example of it not working. If anything, it is any example of it working. Because Protectionism is about protecting your industry from foreign competition - exactly what China decided to do.
There seems to be a belief that the post-war 20th century order would be a fixed feature rather than a contingency of history. That the US would be #1, Europe #2, and the rest of the world would remain "developing" in rural poverty forever. That somehow China could be prevented from catching up. Now they are, and behind them India. I think people are going to be even more surprised in the latter half of the 21st century when South America and Africa start catching up as well.
Brazil is one to watch if they manage to achieve political stability, as is Nigeria if it can transition from being a petrostate.
I’m under the impression that industrial policy has been pretty good for some Asian companies but there have also been notorious failures like the Jones Act.
So I think it can be summarized as “it all depends” and “skill issue.”
> Compared with our initial baseline on the same hardware, we achieved a 3× improvement in end-to-end serving performance, reaching hardware efficiency and per-token cost comparable to mainstream NVIDIA GPUs. This demonstrates that Chinese chips can support frontier-model inference efficiently and economically at scale.
you can also derive some stats from the ~10T tokens a day on 100k devices, 100M / device / day, but then one has to account for the multi-gpu model size, and I need coffee before I go there
I’ll just note that NVidia’s moat has never been inference, and there are many chips used at a much larger scale than Chinese chips for inference like TPUs and AMD chips.
The other part is that it’s a bit of a meme here to say that the chip restriction is actually helping China (or shall I say, coordinated effort?). For once, we know that China has put a lot of pressure on the US to relax these controls multiple times. In addition to large chip smuggling networks (e.g. 22% of NVidia’s worldwide revenue magically comes from Singapore, and the ratio has been growing).
Lastly, assuming acceleration in AI (which we ARE seeing), there might not be time to China to catch up. The best estimate right now is that the first EUV chips from China will not come out before 2030. By that time who knows how powerful AI will be.
All I’m saying is that the discussion is so one sided and a bit baselesss with no nuance, that it seems either a meme/groupthink in the community or coordinated. If anything, the data suggests that the US should increase its export controls and better track the tech supply chain if it wants to further curb Chinese progress.
When you say AI will be powerful, what do you mean? Like in terms of national power, will AI let the US fight and win a war against China despite China's size, manufacturing base and possession of nuclear weapons? Will AI let the US beat the Chinese in manufacturing costs and scale despite China's greater adoption of industrial robots and larger number of skilled workers? Or will AI just be really good at writing software?
Because if the AI is just really good at writing software, I am not sure why China has to "catch up". Seems to me China can just treat AI like other technologies. Let the US pay most of the R&D costs, then come after and treat the technology as a commodity. Sell it better and cheaper and at scale.
I am not sure why it's a big deal for China if China is a few months behind the US in terms of the very frontier AI. Just like I'm not sure if it's a big deal which country has the biggest super computer in the world.
Third party here, I feel like I'm missing something. Your comment seems somewhat pointed and I don't understand what exactly you're asking for?
Weren't they just saying that it's self-evident that preventing a manufacturing superpower - one with a significant pool of industrial engineers and effectively unlimited money & government backing - from acquiring some good is a temporary measure? Because if they have sufficient incentive, they would just... Learn to build it themselves?
After all, their entire nation is built around building things, and catching up / leap-frogging is much easier than starting from scratch.
Not sure, but there are definitely people around the president on both sides, some who think they can addict the Chinese to our silicon, like its the new opium war or something
Yes it was a dumb thing to say and it doesn’t follow at all that restrictions on chips has “sped up” anything. China was always going to build their own chips and their pace is unrelated to a lack of nvidia chips.
Yup. That was really short sighted. And good for China. And actually the overall global market market since supply will augment and competition will decrease pricing as well.
Losers: Chinese AI labs, Nvidia, AMD, TSMC, Micron, SK Hynix, Samsung, Intel.
Any company that depends on Nvidia hardware such as OpenAI, Anthropic, AWS are winners. It means less competition for Nvidia chips and services. If you think Nvidia chips are expensive now, imagine if Chinese companies can buy them freely. Also for American AI labs, it also means they can stay ahead of Chinese AI labs in compute capacity.
The American hardware makers lost the lobby fight in Washington.
In the short term maybe yes, in the long term, maybe they are the winners, they can build on top of cheap inference stack and eventually win on pricing
Such a great move by the government. Imagine what we could do for US industry if the government took a stake in all major corporations? Many people are saying that the healthcare industry should be next.
Where the US sees itself penalizing China with an export restriction, China sees the US gifting it with zero-political-cost “protective” import tariff.
You can’t really hurt a country that has a culture with a positive attitude toward growth.
Nobody pretended that Chinese firms would just lay back and twiddle their thumbs when faced with import restrictions. The question was whether they would be far enough to be able to catch up without much issue or so far behind that they wouldn’t ever effectively catch up or that by the time they did, it wouldn’t matter.
Half-arsed export restrictions are the best of both worlds for these firms: enough of an incentive to take homegrown hardware seriously, yet not aggressive enough to cause meaningful handicap in the meantime.
If someone is capable of doing something, and your goal is to prevent them for doing it, the worst thing you can do is to make it necessary for them to do it.
I might be missing something but when I went to their site they are more expensive than Claude. Why would I pick GLM over Claude? Is it they just offer more tokens in their plans?
For one you would have to use Claude if you pick it. But seriously there is no way for you to determine if one is a better offer than the other, when the usage/tokens/credits are vague, detached, and won't tell you much without trying both.
What are you referring to? Given the audience, my instinct is to assume "plan" refers to the GLM Coding Plans, which are all cheaper than their Anthropic counterparts. As far as I can tell, the API costs are also all cheaper than their roughly equivalently capable Anthropic models.
Not to digress from the core argument of Claude vs GLM being open weights….
I have both plans. Claude monthly €20 and Z’s €18 monthly. Running GLM-5.3 high on their monthly plan will hit quotas absurdly fast compared to Opus 5 High on Claude code. It’s almost unusable for AI driven development. I ended up using the Z plan for using GLM-5.3 as a detailed security reviewer and adversarial feedback. For that, it is much better than Opus which will flag and bail out for even simple security tasks that are aimed at defense.
Just use GLM-5.3 Flash via OpenRouter. It's dirt cheap especially relative to how capable it is. While the Z.ai coding plan was a decent deal in the past I always ran into limiting with it and since I use it intermittently for personal projects my usage wasn't always enough to make the math work - the a la carte pricing via OpenRouter makes this a non-issue.
There's also a new free stealth model available that's more likely than not in the GLM family. This seems to happen every few months for a week or two and represents a good savings opportunity.
This is a really funny sounding post. They sound like they just found out that increasing your automation gives you increased capabilities at faster speeds. They also sound like they just realized AI makes hard things easier.
But what really kills me is the idea that these companies are using Python for production inference. I mean really? Have you seen how bloated and slow Python is? Do global locks really sound like a strategy for fast dynamic computation?
It's not that they "just found out" - what they are saying is that while they were previously dogfooding because it's good practice, now that their models are so much stronger they are using them because it helps accelerate.
If you look at how many years the whole NVIDIA and CUDA ecosystem has been evolving, it's certainly impressive how they've just stood up and optimized this CUDA-free 100,000 node cluster in just a few months.
Most of the fastest inference and training code in production today is written in Python. There are no global locks on the GPU except the ones you put there
If only this infrastructure could handle all the traffic. I've tried using glm via z.ai - and it's a snail kind of slow.
And at the same time you have pretty strict limits to your usage, so in many cases you can't even let it work all night, as you will reach your limit faster than that.
That it's slow doesn't mean it can't handle the traffic, just that this speed is the optimal tradeoff to them. They benefit from serving more tokens by exploiting parallelism across users at a lower number of tokens per second per user, instead of serving each individual user as quickly as possible. When there's a drop in traffic, they probably shut down GPUs rather than giving you higher speed.
Given the huge amount of money being spent on AI chips in the US, what prevents US AI labs from doing the same level of software optimization? It could be a solve for some of the capacity constraints.
They do, when Luna got 5x cheaper it was directly attributed to some unknown % inference optimization.
US labs are quite cut throat about dealing with stuff costing them money (inference). This sort of engineering excellence doesn't always feel that way because they are simultaneously quite lax about stuff costing other people money.
I'm not an experienced engineer, and I work at a small company. To you systems engineers working at large companies, how do you handle these kinds of issues?
- How do you develop new tests and metrics to capture the distinctions you want? Do you work in industries where the abstractions have mostly stabilized? Or is the work in coming up with the measurements themselves?
- How do you stable-ly solve the causality issue? You can prove for one commit in time that some variable was causing an issue by changing it - fine. But that may not address deeper design issues that may not be expressed by that one implementation, if that makes sense. It seems like there's always a meta-level you can go to; sometimes justifiably, sometimes unjustifiably. I liken this to solving individual memory leaks - that's "causal", you can prove that yes, this line of code was the cause - but the deeper issue may be that you're using a memory unsafe language in the first place, and in some sense that's "outside" your ontology. Sure, in the case of memory safety, we've thought about this for decades, so in some sense that's stable enough where it's outside, but known; but what happens when it's outside, but unknown?
I suppose at some point this delves into, "What is good software engineering" in general. And that's not to mention the integration and legacy effects - maybe there's some "fundamentally new better way" to do something, but that requires changing the entire product and a ton of constraints.
And if the response to this is, "Aha! That's the hard part about software engineering!", then how can I get experience in this kind of thing? I know it's abstract but at my current company I haven't developed anything that's lasted more than half a year. I'm wondering if at some point I'd be better of trying to bootstrap experience off of open source.
Maybe now we can stop posting the nonsense take that the frontier labs have hit a wall and are trying to distract from that for IPO reasons.
Also maybe we can stop saying "we can't slow down because China will never slow down" - I don't really think slowing down is right, BUT if slowing down is correct then maybe we should be talking about China slowing down instead of just saying "won't happen" without any evidence that Chinese labs don't have similar concerns.
I have a similar approach where I optimize kernels and find numerical differences between the CPU oracle and CUDA kernels using an automated AI agent in a feedback loop. Usually it solves numerical problems easily (it compares outputs of every layer and compares where they diverge), but so far no matter how many different SOTA models I throw at it, and even show it reference code from other inference engines, they aren't able to much the speed (my engine has a modification which is not found in reference code, although a lot of stuff is similar). Either I'm doing something wrong, or z.ai's Infra Agent is actually an agent swarm, i.e. a bruteforce with heuristics. My project is 2 weeks old so maybe I need more time.
The first part of the article reads like Z.AI is trying to get their piece of the “national security concern” pie.
The way these “AI is too powerful now” articles read about Mythos, Fable, GLM, etc is completely incongruent with my experience using them. It feels like they are all trying to position themselves to influence government policy.
170 comments
[ 6.0 ms ] story [ 79.0 ms ] threadchina has cheap abundant power, now they can make their own inference chips (which was supposed to be a chokepoint), their models yeah can be 6 months behind the frontier - but most people don't need frontier models - small models r more than enough.
my only wish was labs like Mistral would make their own inference chips or partner up eg with established / new chip makers or companies like Oxide.
Only the US sees this as a competition, new space race, Cold War, etc.
Also, don’t underestimate data retention and such. Big Corp will never send their LLM traffic to China.
They must have hit really hard scaling limits if the prices were hiked so much so quickly.
For [API usage](https://openrouter.ai/z-ai/glm-5.3-flash#providers) they charge a bit more than the very cheapest providers of GLM-5.3-Flash, but not so much that a big price difference would make sense.
Can I ask where are you using all those tokens?
I now exclusively use https://omp.sh/ as my harness:
I set it up so it never works in the main branch so subagents etc don't step on each others toes, and only merges back when complete: https://github.com/Daviey/mario/blob/main/.omp/hooks/pre/wor...
A good AGENTS.md is essential: https://github.com/Daviey/mario/blob/main/AGENTS.md
I then provide specifications for what I want, making sure it is unit tested.
https://docs.z.ai/legal-agreement/privacy-policy
Very good - but I'm on a legacy plan. And coming up on a renewal that would put me on the watered down current plan. But with 50% legacy discount think it may be worthwhile. If I go to a competitor I'd be paying market rate.
>They must have hit really hard scaling limits if the prices were hiked so much so quickly.
Not really scaling - their plans were initially comically subsidized even more so than what the western providers are doing. More advert for an upstart than commercially priced.
The max plan will provide ~1,100 USD of GLM-5.3 or ~260 USD of GLM-5.3-flash per month for 168 USD. I can personally attest to these numbers through omp (~97% cache hit rate).
Unless you are able to highly parallelize (your work, you won't be able to hit your hourly or weekly quota using the flash model simply because it's so slow.
They give you ~3x more flash tokens, which maybe comes out to ~2x more actual work after accounting for the extra thinking it does to achieve the same result. The mental model, for not getting angry, is 5.3 is fast mode by default, and you can disable fast mode for 2x the work output at 1/3-1/10th the speed.
They're serving me 5.3 at ~40 tok/s and 5.3-flash at 30 tok/s (according to omp).
That is shocking. Is it per-token I wonder?
I’m getting 97%.
Just checking now: recent runs tau3[1] was at 96% and toolathlon[2] was at 90%
[1] https://www.induction.ai/docs/benchmarks/tau3 [2] https://www.induction.ai/docs/benchmarks/toolathlon
In the PRC, they[1] leaked tons of national secrets on the PRC's latest AI campaigns, the inner workings of their "opinion monitoring" (read: performative panopticon) and "stability" (read: violent oppression) departments, Chengdu's whole CCTV network, direct-energy weapons plans, espionage activities in Syria to hunt down Uyghur refugees, and god knows what else that Anthropic didn't divulge to us common folk.
In the US, it's very clearly an attempt to rip off a competitor. I'm not sure how else you could possibly see it. Even if you're a distillation fan in general (which A. why and B. plz don't), they did this through a network of Japanese and Signaporean shell accounts, presumably at least some of which were abusing Anthropic's subscription service in a ToS double-whammy, as it would be exorbitantly expensive otherwise. They also had to hack around Anthropic's API to get CoT traces, which seems impossible to explain away as anything innocent.
I've been beating the "China isn't necessarily an enemy, it's gonna take us all to handle AI" drum for literally years, but this attack was just... gross. Gross in scale and gross in arrogance. Not a good sign for the dawning alignment crisis, to say the least :(
TL;DR: Use these services if you want, but know that you're supporting aggressive escalations and companies that very clearly don't give a flying fuck about violating the law, much less your ToS. So... buyer beware, I guess.
[1]: For clarity, Z.ai was not alone in this, nor were they most egregious attack -- Moonshot.ai (kimi) took that coveted prize. DeepSeek was involved, too.
Alignment is meaningless; as you've noticed, humans aren't all that "morally aligned".
If the tool needs safety measures it should be kept in a safe enclosure like we do with CNC machines, furnaces, and so on.
I mean, if they get to distill other's IP, why can't others distill their IP?
The Antrophic article mentions "16 million" conversations, GLM models are in the 700-300 billion parameter ranges and while the frontier sizes aren't know but Gemini suggests Astra and Mythos are at around 10 trillion. That'd amount to extracting 40k parameters per conversation without a lot of errors if it was just a distillation (from an unknown source/algorithm as opposed to distilling your own model).
Now, I can imagine these conversations being used as a verification step that they're not missing stuff in their training, and that their models are capable of most of the same things, but that's mostly confirming that they've stolen the same data from the public as Antrophic/OpenAI has stolen already.
Or am I missing something here that makes real "distillation" feasible?
https://x.com/EricSimons/status/2099252922098061714
No real reason to respect any terms they might want to impose. Besides, if you want to break TOS, just have an agent do it; "everyone" running these things agrees there's no corporate or moral liability for what your AI does.
I wonder how you imagine that China built their own space station? Reliant on using American made duct tape, perhaps?
Do you realize how reasoning models are being trained nowadays? You design/build simulation environments to run agents in, with the environment providing the RLVR "verification" scoring. So why won't Ziphu use GLM to build their own RL training environments? Do you think they are not doing this?
Statements dreamed up by the utterly deranged.
Ironically, many benchmarks being maxxed out, and quite quickly, so new ones have to be created.
How, exactly, does "it was Claude that solved Navier Stokes, not ChatGPT!" get you to "AI has hit a wall and stalled"? That's, not to put too fine a point on it, incoherent, and is just noise thrown into the discussion to avoid grappling with the fact that AI continues to rapidly improve.
Some call this "The singularity" (e.g. Hinton).
This is actually a core danger postulated by the, let's call it, "worrying" scenario - see AI 2027 (to be clear, I think its timeline is not realistic).
> Statements dreamed up by the utterly deranged.
Evidently, and tragically, it will take catastrophes to show that deranged are the ones deriding the worried crowd.
Everything you think you are is just what you can imagine in a single moment. Nothing more, nothing less.
Honestly, I have no idea what z.ai is either (I'm aware of an AI-enabled editor called Zed, but that's under zed.dev), so it's a bit presumptuous from them to assume that everyone is familiar with their product...
https://artificialanalysis.ai/#intelligence-category-tabs
Appreciate the clarification. For me it was the "F" in "WTF" that tipped me. Other than that, it's more than fair for you to not know what GLM is. Things are moving so fast that I would be surprised if anyone can keep track of it all. Cheers, have a grand day!
Why would you be reading their corporate blog posts if you don't even know who they are?!
One drive gives about 2 tok/s; striped across four drives it reaches 3.5 tok/s with byte-identical output, and our best internal build with a not-yet-published patch does 4.2.
Method and numbers: https://github.com/argonautlabsai/argodrive (built on antirez/ds4).
GLM has in the past been more technical rather than speculation about future development on RSI etc.
Also curious whether those 100k accelerators are entirely locally made. If that's genuinely end to end on all components including lithography, memory, design etc then that is quite a feat.
Just like the rest of the world, including the US (Intel, Micron), SMIC are currently using ASML lithography equipment (DUV, not EUV), but Shanghai Aishengna are now moving into early production with their own DUV machines, with SMIC and CXMT as early customers.
There is also a state sponsored Chinese EUV development underway.
All I can recall reading from OpenAI about what they have actually done in the name of "RSI" is using one of their models to help automate the training process.
That's more than Anthropic has done though.
In addition to Huawei who make the Ascend series that Ziphu are using, there are also at least a half dozen or so other Chinese companies also making their own AI accelerators.
Seems to work for them.
Almost everyone knew that these sanctions would backfire within a few years. You can't really put sanctions that have noticeable negative effects on bigger economies. They only work for small to medium economies. I believe sanctions on any economy in top 10 would fail.
Nvidia are likely more concerned about AMD taking market share, and I suspect that geopolitics will leave US/China GPUs with largely non overlpping customer bases.
That's not what's happening in the auto-industry, Chinese EVs are easily outselling American ones. Why would GPUs be any different?
This is another example where we might ask "why would it be any different?"
It'd be ironic if Europe ended up banning ASML (a Dutch company) from shipping to US instead!
Mark Carney said earlier today at the EU, "the goal is not self-sufficiency, but collective resilience." It is through collective action that they can find more resilience to our American Antics.
So until China solves the ASML problem, there won't be any flooding.
A bit late for that now that they are moving into early production with their own.
Which they are in progress on: https://www.reuters.com/world/china/how-china-built-its-manh...
China have been really squeezing all they can out of DUV machines, but I'm not sure how much node size really matters for AI competition - more of a cost issue (more chips/power for same FLOPs) than anything, and TSMC & NVIDIA's healthy profit margins need to be considered too.
What is the timeline like? 1 year? 5 years? 10 years?
Small numbers perhaps, but this is happening right now.
What will the numbers be in 5-10 years time? Who knows, but ASML took about 5 years to go from 20/yr to 100+/yr.
Of course politically the world may well be quite different in 5 years time, as may be the AI market.
We also don't know how small of nm chips they can manufacture. If it's 20nm, it's practically useless for advanced AI chips.
Now, the US is left out in the cold with little influence left, themselves now the ones with an anti-missile shortage.
You seem to misunderstand what Protectionism is. This is not an example of it not working. If anything, it is any example of it working. Because Protectionism is about protecting your industry from foreign competition - exactly what China decided to do.
Brazil is one to watch if they manage to achieve political stability, as is Nigeria if it can transition from being a petrostate.
So I think it can be summarized as “it all depends” and “skill issue.”
> Compared with our initial baseline on the same hardware, we achieved a 3× improvement in end-to-end serving performance, reaching hardware efficiency and per-token cost comparable to mainstream NVIDIA GPUs. This demonstrates that Chinese chips can support frontier-model inference efficiently and economically at scale.
The other part is that it’s a bit of a meme here to say that the chip restriction is actually helping China (or shall I say, coordinated effort?). For once, we know that China has put a lot of pressure on the US to relax these controls multiple times. In addition to large chip smuggling networks (e.g. 22% of NVidia’s worldwide revenue magically comes from Singapore, and the ratio has been growing).
Lastly, assuming acceleration in AI (which we ARE seeing), there might not be time to China to catch up. The best estimate right now is that the first EUV chips from China will not come out before 2030. By that time who knows how powerful AI will be.
All I’m saying is that the discussion is so one sided and a bit baselesss with no nuance, that it seems either a meme/groupthink in the community or coordinated. If anything, the data suggests that the US should increase its export controls and better track the tech supply chain if it wants to further curb Chinese progress.
Because if the AI is just really good at writing software, I am not sure why China has to "catch up". Seems to me China can just treat AI like other technologies. Let the US pay most of the R&D costs, then come after and treat the technology as a commodity. Sell it better and cheaper and at scale.
I am not sure why it's a big deal for China if China is a few months behind the US in terms of the very frontier AI. Just like I'm not sure if it's a big deal which country has the biggest super computer in the world.
ok, show it
i`ve recorded a few people who saw this coming in 2022 and i can tell it wasnt many
Weren't they just saying that it's self-evident that preventing a manufacturing superpower - one with a significant pool of industrial engineers and effectively unlimited money & government backing - from acquiring some good is a temporary measure? Because if they have sufficient incentive, they would just... Learn to build it themselves?
After all, their entire nation is built around building things, and catching up / leap-frogging is much easier than starting from scratch.
Winners: Huawei, SMIC, CXMT,Chinese ASML-competitors, OpenAI, Anthropic, Amazon, Microsoft, Google, Meta.
Losers: Chinese AI labs, Nvidia, AMD, TSMC, Micron, SK Hynix, Samsung, Intel.
Any company that depends on Nvidia hardware such as OpenAI, Anthropic, AWS are winners. It means less competition for Nvidia chips and services. If you think Nvidia chips are expensive now, imagine if Chinese companies can buy them freely. Also for American AI labs, it also means they can stay ahead of Chinese AI labs in compute capacity.
The American hardware makers lost the lobby fight in Washington.
In the short term maybe yes, in the long term, maybe they are the winners, they can build on top of cheap inference stack and eventually win on pricing
I don't think your information is entirely accurate.
Just logic.
You can’t really hurt a country that has a culture with a positive attitude toward growth.
Half-arsed export restrictions are the best of both worlds for these firms: enough of an incentive to take homegrown hardware seriously, yet not aggressive enough to cause meaningful handicap in the meantime.
What's the alternative other than a military/naval embargo?
If someone is capable of doing something, and your goal is to prevent them for doing it, the worst thing you can do is to make it necessary for them to do it.
GLM: 80 USD (pro), 168 USD (max) -> with "limited-time event" discount this becomes 56 USD and 117.6 USD
I also don't understand why are they so much costlier, and I would also like to give it a try.
GLM's "Max" plan is (was?) equivalent to 3x Claude's 20x ($200) plan.
Anthropic's Pro is $20 and corresponds to Z.ai's Lite at $18
Anthropic's 5x Max is $100 and corresponds to Z.ai's Pro at $80
Anthropic's 20x Max is $200 and corresponds to Z.ai's Max at $168
I have both plans. Claude monthly €20 and Z’s €18 monthly. Running GLM-5.3 high on their monthly plan will hit quotas absurdly fast compared to Opus 5 High on Claude code. It’s almost unusable for AI driven development. I ended up using the Z plan for using GLM-5.3 as a detailed security reviewer and adversarial feedback. For that, it is much better than Opus which will flag and bail out for even simple security tasks that are aimed at defense.
There's also a new free stealth model available that's more likely than not in the GLM family. This seems to happen every few months for a week or two and represents a good savings opportunity.
But what really kills me is the idea that these companies are using Python for production inference. I mean really? Have you seen how bloated and slow Python is? Do global locks really sound like a strategy for fast dynamic computation?
If you look at how many years the whole NVIDIA and CUDA ecosystem has been evolving, it's certainly impressive how they've just stood up and optimized this CUDA-free 100,000 node cluster in just a few months.
And at the same time you have pretty strict limits to your usage, so in many cases you can't even let it work all night, as you will reach your limit faster than that.
US labs are quite cut throat about dealing with stuff costing them money (inference). This sort of engineering excellence doesn't always feel that way because they are simultaneously quite lax about stuff costing other people money.
Signed, a customer.
- How do you develop new tests and metrics to capture the distinctions you want? Do you work in industries where the abstractions have mostly stabilized? Or is the work in coming up with the measurements themselves?
- How do you stable-ly solve the causality issue? You can prove for one commit in time that some variable was causing an issue by changing it - fine. But that may not address deeper design issues that may not be expressed by that one implementation, if that makes sense. It seems like there's always a meta-level you can go to; sometimes justifiably, sometimes unjustifiably. I liken this to solving individual memory leaks - that's "causal", you can prove that yes, this line of code was the cause - but the deeper issue may be that you're using a memory unsafe language in the first place, and in some sense that's "outside" your ontology. Sure, in the case of memory safety, we've thought about this for decades, so in some sense that's stable enough where it's outside, but known; but what happens when it's outside, but unknown?
I suppose at some point this delves into, "What is good software engineering" in general. And that's not to mention the integration and legacy effects - maybe there's some "fundamentally new better way" to do something, but that requires changing the entire product and a ton of constraints.
And if the response to this is, "Aha! That's the hard part about software engineering!", then how can I get experience in this kind of thing? I know it's abstract but at my current company I haven't developed anything that's lasted more than half a year. I'm wondering if at some point I'd be better of trying to bootstrap experience off of open source.
Also maybe we can stop saying "we can't slow down because China will never slow down" - I don't really think slowing down is right, BUT if slowing down is correct then maybe we should be talking about China slowing down instead of just saying "won't happen" without any evidence that Chinese labs don't have similar concerns.
The way these “AI is too powerful now” articles read about Mythos, Fable, GLM, etc is completely incongruent with my experience using them. It feels like they are all trying to position themselves to influence government policy.