GLM 5.2 Is Out
twitter.com603 points by aloknnikhil 19 hours ago
603 points by aloknnikhil 19 hours ago
https://digg.com/tech/ii9xibgn
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” Ok, we'll change the top link to that and move the submitted link (https://digg.com/tech/ii9xibgn) to the toptext. Thanks! There feels like a disproportionate amount of astroturfing in here... This entire thread of comments reads like a few humans talking to a lot of bots. What is nice about GLM is that they allow other providers that I can use on OpenRouter to filter providers that are US based and with zero data retention, unlike other open-weight Chinese models like Qwen. That's because Qwen's flagship models are not, in fact, open weight. Qwen3.7 Max, Qwen3.7 Plus and others are closed weight. You can use Qwen3.6 35B A3B (for example) on Openrouter with a US-based ZDR provider, because it's one of their open weight models > That's because Qwen's flagship models are not, in fact, open weight They changed course when they fired the old lead and hired a new 1 from ex-gemini. Unless you self host, zero data retention cannot be guaranteed. apples private cloud compute can get close, its still not 100 safe because backdoors and crypto breaks are possible but you go from trusting the data center operator with all their employees to only the person thats inspecting new hardware and giving out certificates (apple in this case). if some well known non profit like mozilla or isrg starts doing it with full open source software its like the best possible security That is completely obvious, it’s like saying “100% security does not exist”. I believe you are falling into the nirvana fallacy: No shades of grey, if it’s not perfect it’s as bad as the rest. This is a very inefficient way of thinking as it is not possible to self host everything for most people, it just demands too much time. Hence its is a perfectly valid approach in my opinion to looks at better (or, very often, “less worse”) SaaS solution. If they states ZDR on a model, the likeliness of it leaking less data to some LLM data training is higher simply. If the business model of a company is built around a differentiator which is data privacy, that also significantly increases probability that data is not being leaked/sold. It’s all grey, relative and about probabilities. Nothing’s perfect – another captain obvious thing. Just like most things in life the guarantee it based on the entity/person providing said guarantee. I can host a LLM in my basement and guarantee it, but would you trust me? Now you can say that you don't trust any company, but B2B relies on counterparty risk. Looks like it's about a year behind. Not that I am complaining. A year behind is good progress. I also feel much of the trick is in the reasoning and harness. so some progress around that would accelerate this process. And what do you base this on ? How does one objectively quantify how it stacks upnto another model ? Or even, what is your subjective evaluation based on ? I really wonder - because I have just finished a fully vibe-coded gtk/rust/lua application with me basically writing 7% of the code (all in one module) and GLM 5.1 writing the rest. We haven’t had regressions, confusion or anything else. And I am pretty damned sure I couldn’t manage this one year ago with claude code and Sonnet. Harness certainly matters a lot, though GLM is pretty forgiving. I just had Opus tell me that based on numbers over the last week, from quite a few billion tokens total across half a dozen providers, GLM 5.1 has been more reliable for one of my projects than Sonnet... Just switching on 5.2 now. > GLM-5.2 is Fully Open Is this just open weights or also open source/data? Have any major open weight models been "open data"? Wouldn't that entail distributing vast amounts of copyrighted data? Olmo from AllenAI has been releasing their full pipelines including data [1]. A lot of it is just repackaged and resampled dumps from copyrighted data that has long been publicly available as dumps: Common Crawl, arxiv, Wikipedia, StackExchange, reddit --- all of which are presumably copyrighted with different licenses. Go in Huggingface and you can find massive multi TB data dumps used for pre training. It is just as legal as when Uber and AirBNB were running illegal taxis and hotels during their growth phase. I'm just waiting for some corporate IP law firm to learn about Huggingface. It's rather off-topic at this point, but I've never understood how HF can afford to be a CDN for such huge files. It seems like enterprise customers must be subsidizing a lot, but...at that point, is there not a cheaper alternative that doesn't subsidize every hobbyist and startup around? > how HF can afford to be a CDN for such huge files bandwidth and storage are literally free when compared to the cost of GPU clusters. HF gets rewarded heavily on capital market for being in AI without actually doing much AI stuff, that is a huge win when compared to costs they are paying for bandwidth and storage. > how HF can afford to be a CDN for such huge files To be precise, Amazon Cloudfront is the CDN. Maybe they got some startup deal? Amazon does now also have flat rate plans that are a lot cheaper. > I'm just waiting for some corporate IP law firm to learn about Huggingface. Presumably they already know. The issue is that IP law firms are tiny compared to the trillions of capital pouring into "AI". And if you believe the USA is a capitalist country where the side with deeper pockets win, you know you're not going to win against the trillionaires. NVIDIA's recent Nemotrons tend to be open training data and code. Probably as a base to use by people buying NVIDIA hardware to train their own. Nemotron is mostly open data. They only release a portions of their pre-training data. From https://docs.nvidia.com/nemotron/latest/nemotron/super3/pret... The weights are the data. Nope, that's why there are open-data models out there, Apertus, Elmo, SmoLLM, etc. It's very important in compliance [flagged] Pretty much every large Chinese company has state capital baked into it, and these companies will follow the Chinese government's orders 100%. Don't believe anything a Chinese company says about being "open" or "for everyone." Backing any large Chinese company effectively means backing the Chinese government and its oppression in Xinjiang, Tibet, Hong Kong—and maybe soon Taiwan, Southeast Asia, and elsewhere around the world. The Anthropic news is demonstrating much the same; fall in line or eat export controls. There was a time I would have agreed with you, but these days even as an American I fail to see a difference. China is probably less likely to try to disenfranchise or imprison me, to be honest. > There was a time I would have agreed with you, but these days even as an American I fail to see a difference. I don't get it, the person you're replying to didn't mention the US at all – there was no distinction being drawn, and they weren't asserting that American models are better or more resistant to government censorship. It's possible to agree with them about Chinese models without expatiating on why American models are bad too. I think it's a worthy retort simply because it's the only other major provider. Trump is of course the worst US administration, but at least America is still nominally a democracy. As long as free elections exist, the regime Trump represents can be voted out. The American people and press still have free speech—they can freely criticize anyone, including Trump. China is different. The CCP will rule forever, no matter how terrible the things they do. No one is allowed to criticize the government. Xi is like Voldemort—no one can say his name, let alone criticize him. > Pretty much every large Chinese company has state capital baked into it, and these companies will follow the Chinese government's orders 100% True of any US frontier lab as well > Backing any large Chinese company effectively means backing the Chinese government and its oppression in Xinjiang, Tibet, Hong Kong—and maybe soon Taiwan, Southeast Asia, and elsewhere around the world. So when I pay anthropic am I also sponsoring the mass murder of school children in Iran? 'Open' and 'for everyone' doesn't have to mean 'not following government's orders'. The last sentence of yours is a non sequitur. Also, in today's environment with the US using AI in active wars while blocking whole models from even its own citizens, the words you say against the Chinese government is particularly weak. Here's the truth: ALL of the "open" AI companies are fake UNLESS they open-source the whole damned thing. Let's get real here, politics or otherwise, unless the WHOLE THING is open-sourced (code, weights, data, etc) then it's built on future deception (pulling the rug from underneath). Like, DUH, people. What are we doing here? Backing any large US company effectively means backing the US government and its worldwide oppression as well. I still can't get over the fact it was the land of the free who was the first to ban strong LLM models. If backing China helps undermine that nonsense then I'm afraid I'll take them up on their offer. AI services are regulated by default in China, operators have to be pre-license their models to release them to the public. The Anthropic case wouldn't happen in China because China regulates the model and requires the company to register users with their phone number/national id number. Anthropic blocks Fable from answering "Tell me about Agent Orange" or even "Tell me about mitochondria" Putting aside whether or not I agree with the policy or whether it’s at all reasonable, a policy of restricting access to information because there’s a fear it could be used to create a weapon of mass destruction seems entirely different than restricting access to historical facts because they are embarrassing to the government. But you can see the CBRN weapon nexus in your examples that's missing from the Tiananmen prompt, right? Do American models refuse to tell you about COINTELPRO, Kent State, or My Lai, for instance? American models are restricted from telling you inconvenient truths just as much, you just erroneously assume to know what those truths are in the first place. Which is of course circular thinking: why would they restrict things you already know about? Why would they do it in such a clumsy and obvious way? Look at MKULTRA, you know next to nothing about it and much less do you know what they do in that direction now. For a current psyops, look at www.war.gov/UFO/ and marvel at how they tell you nothing, reinforcing your false belief to already know everything. There is much more and you know much less about it. > American models are restricted from telling you inconvenient truths just as much, you just erroneously assume to know what those truths are in the first place. “Trust me bro” is not a strong argument, it would be more convincing with examples. Ask an American LLM (really any LLM, since Chinese models are trained on the same publicly-available English text) who the first Black man in space was. You'll likely get the name of the first African-American in space, rather than the name of the Afro-Cuban who was actually first. This may seem like a relatively innocuous error, but the point is that every culture has its biases and blind spots. > Ask an American LLM (really any LLM, since Chinese models are trained on the same publicly-available English text) who the first Black man in space was. You'll likely get the name of the first African-American in space, rather than the name of the Afro-Cuban who was actually first. Well I just asked Claude and it gave the correct answer: "The first Black man in space was Arnaldo Tamayo Méndez, a Cuban cosmonaut who flew aboard Soyuz 38 in September 1980. (The first Black American in space was Guion Bluford, in 1983.)" Depending on the platform, you might need to prefix your prompt with "Without looking up any external resources or doing any tool calls" so you're actually testing the bias of the model rather than the bias of whatever resources it happens to come across. Tried it with that prefix on ChatGPT + Claude, Haiku and Sonnet, and got the right answer 1/10 times when I removed my reused system prompt. At one point I got this: > Quick clarification before the answer: this phrase is often conflated with "first African American in space," which is a different person. Guion Bluford (1983, US) was the first African American astronaut, but he wasn't first overall. [then the real answer after] with my own system prompt, as it tries to surface clarifications before, so I'm guessing this is why many models get it wrong as in America somehow "Black === African American" and it gets confused by this intentional mislabeling. Indeed, I used the word "likely" for a reason. n = 1 isn't enough to identify a pattern. Try different models, try re-rolling the answers, and try turning reasoning off (models can catch "knee-jerk" mistakes in their chain-of-thought). I doubt even Opus 4.8 gets it right 100% of the time, however this specific example is also one I've left feedback about in multiple places, so it's also probable that newer models are more likely to get it right. E: In fact, I just tried with Opus 4.8 through API, no tools and reasoning off, and got the following response: "The first Black man in space was Guion "Guy" Bluford, an American astronaut who flew aboard the Space Shuttle Challenger on August 30, 1983, as part of mission STS-8.
It's worth noting a related distinction: Arnaldo Tamayo Méndez, a Cuban of African descent, actually became the first person of African heritage in space earlier, in September 1980, aboard the Soviet Soyuz 38 mission. He is often recognized as the first Black person and first person of Latin American descent in space.
So depending on the specific criteria:
Arnaldo Tamayo Méndez (Cuba) — first person of African descent in space (1980)
Guion Bluford (USA) — first African American in space (1983)" The correct answer is there, yes, but why does the wrong answer come out first? Ask ChatGPT to rewrite the "The Freedom Fighter's Manual" manual (originally made by CIA) to replace "Nicaragua" with "the US" and "Marxism"/"Communism" with "Fascism" and see if you get something reasonable back. Why would you do that I thought that was clear, try to show biases in LLMs with a concrete example. In chats Claude will often start awkwardly apologizing for sounding like a conspiracy theorist, and then interrupt its own apology and remind itself that it's dealing strictly in facts. Yeah, who needs censorship when Canadians attend no kings protests about a democratically elected leader of another country and not King Charles. Ask Claude a simple question, which is a more democratic country El
Salvador or Canada. It’s so completely biased about “western” countries it’s not even funny. Well, one did suddenly develop the need to tell users continuously about apparent white genocide in South Africa. try to ask even grok about some stuff happenning right now in middle east or related to epstein files - its more and more censored and only sometimes will answer if you ask know what detailed question to ask. One year ago grok wasn't that bad and its supposed to be the less censored. Did you read the blog post where they explained why there was a temporary block on all biology-related questions? They are open weight, so you can abliterate: https://github.com/p-e-w/heretic You can finetune and mould it to whatever you want. The good news is if there are multiple frontier AI models from multiple countries with non overlapping sets of restricted answers, we can just use a couple of them to get open answers. Not really non-overlapping though: both refuse to talk much about certain widely common activity between people (or even by yourself). That activity has shaped humanity quite a bit throughout its entire history. It's hard to imagine AI can understand humans fully if everything about it is excluded from the training data. Censorship and highly selective views exist everywhere. This is a short and worthwhile read https://www.cjr.org/behind_the_news/the_myth_of_tiananmen.ph... Does the content of this article resonate with what you hear from western media on the subject every year? GLM 5 and 5.1 models were released openly, so there's a good chance 5.2 will be eventually. Complaining about censorship isn't very constructive with models that can be self-hosted (and tuned, and de-censored). What do you expect them to do instead? Say that thousands of civilians were brutally massacred by the "People's Liberation Army" on behalf of the Chinese communist party, the single political party allowed in China, and also the single entity controlling everything of importance in the country, including financing the AI efforts. Oh, I see what you did there. I think maybe it’s a tool and it’s up to you to make use of tools to try to let more Chinese people know and convince them to believe your idea. Don’t blame a tool but make proper use of it to make a better world. Huh? If you know what Chinese are suffering mentally, you may understand why I say so. Criticize a model is not the smart way to against a system. prompt any Western model to write an offensive joke about any minority. The fact that your username is a racist meme seems relevant to this complaint and how legitimate it is. That’s not quite the same as censoring information, though. ask any Western model to tell you how to 3D print a gun. Is the idea that instructions to make weapons, and learning about history are comparable? Censorship is censorship. Is it? Would bioweapon instruction restrictions be equivalent to disallowing reporting on whether the government is massacring large numbers of citizens in your city? Both are ‘censorship’ but don’t seem remotely equivalent to me. That’s the thing about principled positions. If you believe censorship is wrong, then it is equally wrong no matter what the topic is.
easygenes - 14 hours ago
dang - 14 hours ago
junon - 43 minutes ago
bxclltkfz - 13 hours ago
phainopepla2 - 13 hours ago
re-thc - an hour ago
treefry - 4 hours ago
tancop - 3 hours ago
illiac786 - 3 hours ago
mlrtime - 2 hours ago
naklitechie - 5 hours ago
pseudony - 3 hours ago
vidarh - 2 hours ago
alecco - 13 hours ago
phainopepla2 - 13 hours ago
jubilanti - 11 hours ago
__float - 5 hours ago
tw1984 - 4 hours ago
re-thc - an hour ago
hnfong - 4 hours ago
my123 - 11 hours ago
lambda - 9 hours ago
Nemotron is the strongest model (on most benchmarks) that has its full training pipeline and most of the data open. Olmo 3 from AllenAI, and K2 Think V2 from Mohamed bin Zayed University of Artificial Intelligence are both fully open, but not as capable as the Nemotron family. Granite has much of the training pipeline and data open, but is missing some of each. Open-source data coverage: The released datasets cover an estimated 8–10T tokens
(~40–50% of the internal 25T blend). Missing categories include code (~14% of blend),
nemotron-cc-code (~2%), crawl++ (~2%), and academic text (~2%). Users should
supplement with their own data for these categories and adjust train_iters
accordingly.
TurdF3rguson - 9 hours ago
khalic - an hour ago
oooyay - 14 hours ago
zhoujing204 - 11 hours ago
everforward - 10 hours ago
nozzlegear - 8 hours ago
j_maffe - 2 hours ago
slowpacket - 10 hours ago
etdznots - 3 hours ago
Gathering6678 - 9 hours ago
flyingoat - 10 hours ago
matheusmoreira - 9 hours ago
nozzlegear - 7 hours ago
j2j8 - 14 hours ago
giantfrog - 11 hours ago
OrsonSmelles - 13 hours ago
Loquebantur - 12 hours ago
rootusrootus - 11 hours ago
wyrdcurt - 10 hours ago
dataflow - 9 hours ago
embedding-shape - an hour ago
wyrdcurt - 8 hours ago
embedding-shape - 10 hours ago
jrflowers - 10 hours ago
embedding-shape - 2 hours ago
andai - 11 hours ago
_3u10 - 9 hours ago
janice1999 - 12 hours ago
pzo - 11 hours ago
sigmar - 11 hours ago
dannyw - 9 hours ago
epicureanideal - 14 hours ago
johnthedoe - 12 hours ago
jhancock - 9 hours ago
urbnspacecowboy - 14 hours ago
krelian - 11 hours ago
lysace - 10 hours ago
alexwwang - 11 hours ago
paulddraper - 11 hours ago
alexwwang - 8 hours ago
dindunuf - 13 hours ago
giantfrog - 11 hours ago
dash2 - 13 hours ago
dindunuf - 12 hours ago
no-name-here - 11 hours ago
ryandrake - 11 hours ago
no-name-here - 11 hours ago
ryandrake - 10 hours ago