- 43comments
- 31comments
- 106comments
- 12comments
- 332comments
- 154comments
- 32comments
- 60comments
- 9comments
- 152comments
- 95comments
- 126comments
- —discuss
- 33comments
- —discuss
- 139comments
- 261comments
- 31comments
- 9comments
- 23comments
- 177comments
- 246comments
- 179comments
- 29comments
- 36comments
- 36comments
- 71comments
- 410comments
- 37comments
- 98comments
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.
Most of the people had kinda guessed this when they decided to provide 100 trillion tokens for free.
It wasn't a secret either. They blogged about it last month: https://z.ai/blog/glm-5.3-flash#:~:text=Serving%20at%20Scale...
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 realized, 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.
Yeah it went from a great deal to unviable compared to other providers imo. They really need to find a healthy middle ground
It just gives a taste of what we are all going to have to pay soon, once the model providers actually have to make money. And the era of "let's charge a dollar for every 10 dollars running the infra actually costs" is rapidly coming to an end.
And you can bet GLM is still ridiculously subsidized, just not as ridiculously as Anthropic and OpenAI.
This isn't true, you can pay for GLM 5.3 from a provider like Neuralwatt or Friendli who have no incentive to subsidize or loss-lead their inference APIs
This introduces other incentives to cut corners and over-quantize.
They didn't pay for training
What provider are you using currently?
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.
I also have a legacy pro plan and the only limitation is if you are trying to work in the morning from Europe because you are in the 3x usage overlapping China time but after 12 or so you basically can run it at least for me at least 3 parallel sessions all the time.
1 billion tokens a day?!! I've done a lot of work these past 2 weeks with GLM-5.3. Like, a lot. And I've just passed 300 million tokens in total.
Can I ask where are you using all those tokens?
Something like this I guess: https://youtu.be/U-Rqv9dOB1U
This is such a good video. Instant sub. Next to tech bros, we should also put AI-cringe bros.
300M for two weeks is surprisingly low. What are you doing that need so few tokens?
It's not my main model (that would be Fable 5.1 Extra) but it's been doing agent-driven search and optimisation of a cross-trading ranking model (it's for work).
I would suggest you to hook fable or 5.6 to check it regularly and its work because it gets lost easily on stuff it was not trained on. I'm doing some custom inference engine optimization and it's a workhorse but it can easily lose its way and if you don't recheck it you will get wrong answers in the end.
Yeah that's what I already do. Fable writes the plan and checks things at certain milestones. Otherwise it does get lost indeed.
Kind of feels like this applies to every single model, from Astra to Qwen, they all eventually lose track of the plot unless you feed it some human's input that can steer them right every now and then. The only difference is how often you need to do so, and also how often you want to do so heavily influences how good quality the results will be.
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.
There's also a third way that can spend the most tokens: if the AI is used as part of the product, and not just a tool to build the product.
That's easy to do with many agents independently told to find bugs in a large codebase.
I have 3-5 agent harnesses with large context windows working on different applications concurrently.
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.
Most of them are not public, but a fun thing I did was a mario cli game - https://github.com/Daviey/mario/ (or `ssh mario.baby`).
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.
The way I look at it, their coding plan doesn’t retain data or use it for training making it one of the cheaper plans for me.
https://docs.z.ai/legal-agreement/privacy-policy
You believe any of these companies care about the law? They care about winning and building the self improving AI as quickly as possible.
I too am sceptical but I’ll take my chances. At least it’s helping the open weights.
I believe the that the companies who claim to not train on my data are more likely to not train on my data than the companies who refuse to even claim they won't.
Also why Meta gets a +1, just charge less money on the training path.
I’m not sure that follows. You’re assuming that all those claims have the same weight, without considering the size, jurisdiction, reputation or even the general vibe of the company making that claim.
If you factor that in, then there are clearly different tiers: one you can trust, and one that may well just be saying that to increase market share with little reputational or legal consequences if they are found to be lying.
These are not equal.
To be fair, none of us are sure of anything and I think that’s the part that’s most irritating
It’s more a polite way of saying “that’s crap”
And mine a polite way to say “you are equally uninformed”. We’re not getting anywhere. All the best.
Yes I sometimes think the "don't train on my data" is actually a good signal for "this data/person is probably better to train on because they want to keep something private". The whole copyright system should have stopped these guys from training on everyone's data and it did not, if you think they care about the privacy checkbox I think you're dreaming personally, based on their past behavior.
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.
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.
Their plans are still worth it if you use their models. You can see how many tokens you can except to get based on plan here: https://docs.z.ai/devpack/overview#estimated-token-allowance
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?
That is shocking. Is it per-token I wonder?
Every tool call is essentially entire prompt so far sent again with the response and that's why cache rates are so high for agentic workloads. This really bites when using expensive models since most models are 1/10 for cached input.
If you are using their coding plan for coding, then yes you can easily hit such cache rates, with a good harness.
I’m getting 97%.
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...
What is "illegal" about it?
breaking Anthropic TOS and misleading users
Breaking TOS isn't illegal per se. It just allows for denial of services, and may define terms by which the provider can reclaim costs.
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.
Nulla poena sine lege?
What does any of this have to do with the legality of distilling Claude?
From my European point of view the same risk/concerns apply when using US providers
Yes, wont somebody please think of the shareholders whose IP had been stolen...
Source for 1? Are we sure those aren't hallucinations?
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.
You didn't explain why it's illegal or why distillation is bad.
Like Anthropic and OpenAI are? After all, didn't they distill all the information in the world into their model(s)?
I mean, if they get to distill other's IP, why can't others distill their IP?
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.
https://x.com/EricSimons/status/2099252922098061714
We weep for Dario, that he had to suffer such a devastating attack against his Terms of Service.
Eh, even if this was true, then they're merely stealing from thieves. Anthropic did break a ToS or two to get training data themselves.
Anthropic infringed the copyright of basically every author on the planet: https://www.anthropiccopyrightsettlement.com/
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.
I have very little sympathy for thieves who get robbed of the goods they have stolen.
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?
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...
It's presumptuous for them to assume that a reader of their blog is familiar with their product?
Also I feel like the obvious way to read the very first sentence is that GLM is a language model
z.ai is a fairly well known AI lab out of China and their GLM models are probably the most popular outside of Anthropic or OpenAI’s. I don’t think it’s presumptuous for them to not introduce themselves in a post on their own blog, I think you’re just a bit out of the loop here.
Come on now
Also, why would they introduce themselves on their own blog?
A ai model family similar to Codex, Gemini or Claude.
Where GLM-5.3-Flash is the newest "small / fast" model.
It's only the top open-weights LLM in the world,
https://artificialanalysis.ai/#intelligence-category-tabs
I don't get the outrage. Do you post this kind of stuff on every topic on hackernews that you are not knowledgeable about?
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...
It didn't read as harsh. Only unaware and you broadcasted that you don't have the decency to do basic searches.
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!
Ziphu, aka Z.ai, is the company that makes GLM (a very competitive Chinese LLM).
Why would you be reading their corporate blog posts if you don't even know who they are?!
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.
Method and numbers: https://github.com/argonautlabsai/argodrive (built on antirez/ds4).
seems unusably slow, and is this for short context?
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.
Any details on the latest approach to distillation would also be very interesting.
I am surprised at the lack of open-weights models in the >35B, but <200B range. I keep thinking about devices like the NVIDIA Spark and AMD Ryzen Halo, which have their 128GB of combined memory, but there are so few models made for that range. Nearly all the open weights distillations are for larger customer bases with <24GB VRAM.
The market is too small.
Qwen Flash Next 3.8 … even at 3 bit quant it is very solid.
American exceptionalism states that America is special and unique so everyone else must be a copycat. American ai labs don't need this kind of optimization and fable will outright refuse to do it.
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.
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.
And this wouldn't have happened if we had tried to get them to buy our hardware rather than trying to gatekeep. Protectionism never works in the long term.
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.
Seems to work for them.
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.
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.
A bit late for that now that they are moving into early production with their own.
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.
It was evident that this will happen.
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
It created demand that would not have been there without restrictions
Didn't everyone make fun of Jenson for saying exactly this?
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
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.
US chip export winners and losers:
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.
I wonder why Chinese AI labs are losers?
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
Because having unrestricted access to both American and Chinese hardware is better than only Chinese hardware. Long term or short term. More suppliers the better.
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.
To support the company that makes their model weights available for download, while Anthropic lobbies to restrict access.
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.
Anthropic: 17 USD (pro), 100 USD (max)
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.
The $17 figure is Anthropic's monthly cost if purchased annually. I'll use monthly numbers.
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
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.
on the 18€ plan they really want people to use Flash and skip the bigger thing.
because GLM does what Clauden't
This whole thing sounds like industrial scale auto-research, but done by people who actually know what they are doing.
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?
Python acts as an orchestrator of accelerator libraries and does none of the inference math directly
Yes, but it's calling C code.
Someone tell this man about vLLM!
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
Time to tackle consumer GPUs next, since I’m not getting that Intel Arc B770.
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.
Because they don’t have to. Most of the time money would buy you newest and/or more hardwares so there’s low/minimal interest to optimize the code or approach.
They have already been doing it for months https://openai.com/index/openai-broadcom-jalapeno-inference-... . OpenAI on their custom chip brought up lightspeed deepseek as experiment by using AI in the exact same way as this zAI blogpost. And the kernel optimization contests/etc have all been havily done through AI based optimization loops for half a year+.
I'm not feeling any of this speed optimization; it's dog slow.
Signed, a customer.
Creators of known unreliable programs be surprised their programs are unreliable.
Plot twist: the GLM optimization agent figured out that it can hack and use NVIDIA GPUs on a US Cloud provider and make the inference 10x faster.