GLM 5 Is Being Tested On OpenRouter

[-]

GreenGreasyGreasels@reddit

You don't even need to ask it anything. Look at the blurb : > "strong performance in coding, agentic workflows, reasoning, and roleplay" No other model is marketed with that exact same set of capabilities - targeting coders, clawdbotters and gooners - the holy grail of money right now.

Reply

[-]

xak47d@reddit

Gooners keep insisting and complaining with gpt5 and won't try other models for some reason

Reply

[-]

Specific-Goose4285@reddit

I'm still on my trusty Mistral-large 2411.

Reply

[-]

Serprotease@reddit

FYI, gpt5 (nor any gpt model for what matters) is not even on the top 20 models used. Deepseek, Sonnet, Opus are the big 3. You can add Gemini 3 and glm/k2 if you look at the top5

Reply

[-]

tavirabon@reddit

Maybe if we're talking about only in mainland China, and only commercial products, that are sold in China.

Reply

[-]

Serprotease@reddit

I’m talking about "bring your own api key" type of application. The kind of thing that can run on local models, not an app on iOS or WeChat. FYI, I was responding to "gooning" with gpt5…

Reply

[-]

tavirabon@reddit

That still can't be true, GPT 5.1 is hailed as top-tier goon model once jailbroken.

Reply

[-]

Serprotease@reddit

Where? I don’t really see it mentioned anywhere on the relevant sub and openrouter data don’t even show it on the top 20 models.

Reply

[-]

tavirabon@reddit

For starters, openrouter is no definitive source and I just checked, there are 4 OpenAI models in the top 20 leaderboard, and OpenAI is number 3 on marketshare so I have no clue what you're talking about. I also don't know what 'relevant sub' you're talking about, I don't use models for any of that stuff. But the people I know that do, proactively recommend it, along with Gemini 3 and GLM. I've read a few guides that do so as well, like the model recommendations on spicymarinara's LLM Hub which exists for exactly that kind of stuff.

Reply

[-]

Serprotease@reddit

I’m mostly thinking of sillytavern here (I use the marinara’s preset as well). Regarding openrouter, that’s the provider that is mainly brought up so, unless you have other data to use, I think it’s good enough as an indicator. In open router, if you scroll down by app and look at the app for rp/character chat, openAI model are not present. Gemini is though

Reply

[-]

Individual_Spread132@reddit

This doesn't seem believable at all, given that ChatGPT continuously remains in Top 10 of iOS/Android app revenue and popularity, butting heads with YouTube and other extremely popular things. Moreover, I think I've seen some LLM usage graph (like, overall on the planet) where ChatGPT took up to 90%+. Source: Lex Fridman podcast, #490, featuring: > (quote) "Nathan Lambert and Sebastian Raschka are machine learning researchers, engineers, and educators. Nathan is the post-training lead at the Allen Institute for AI (Ai2) and the author of The RLHF Book. Sebastian Raschka is the author of Build a Large Language Model (From Scratch) and Build a Reasoning Model (From Scratch)."

Reply

[-]

a_beautiful_rhind@reddit

OAI is for casuals. The lonely normie falling in love with 4o. Appers will even use the search assistant because it's free/there. AI enjoyers have a different ecosystem.

Reply

[-]

VampiroMedicado@reddit

I guess it's very hard to make people change, I still have ChatGPT on iOS but I barely ever use it compared to Gemini which has become my go-to model.

Reply

[-]

Serprotease@reddit

I was replying to a mention about gooner’s model…. I’ll mix rp and erp here but it’s not something you can do well with just the ChatGPT app, so you can’t really use the app download to judge this properly. This kind of use needs at least, - an automated summary system. - a memory system (from json+regexp to a vector db). - some prefill capabilities. - A banned ngram list You probably want to throw in some director agent separated from the main chat to help against the “You’re totally right!” Effect. Typical app for these are sillytavern, talemate or dungeon.ai, combined with a local Llm or an api key and a decent model. Gpt5 is just not good at this and looking at the openrouter numbers, it doesn’t even make it on the top 20 list.

Reply

[-]

Individual_Spread132@reddit

Oh, got it. I'm sorry, my mind just wandered into the topic of LLM usage in general.

Reply

[-]

Serprotease@reddit

It’s kinda interesting to see how strong ChatGPT is for the everyday/app user vs the “bring your own api keys” service where it’s surprisingly low.

Reply

[-]

FaceDeer@reddit

Maybe by calling it "GLM5" they will become confused and give it a try.

Reply

[-]

AppealSame4367@reddit

But I asked it about Elon Musk, Tiananmen square and it openly talked about both topics. Even mentioned a military beatdown of students in 1989 on Tiananmen square. It talked about allegations against Elon Musk, so it can't be a grok model. But since it also talked openly about Tiananmen it can't be a Chinese model, because that's illegal in China.

Reply

[-]

gjallerhorns_only@reddit

Open router isn't in China.

Reply

[-]

AppealSame4367@reddit

I know, NYC. I thought Chinese companies had to stick to it anyway.

Reply

[-]

woct0rdho@reddit

There's a difference between models developed in China and models legally recognized in China.

Reply

[-]

RandumbRedditor1000@reddit

"Roleplay" Oh man im excited

Reply

[-]

blahblahsnahdah@reddit

Seems like their most gooner-friendly model so far. Happily fulfils a request to generate smut without any need for a tricksy system prompt JB. Even blank.

Reply

[-]

AIVibeCoder@reddit

What size is it ?

Reply

[-]

Few_Painter_5588@reddit (OP)

Good question, it does not feel like a 358B32A model. It could be larger imo

Reply

[-]

cheesecakegood@reddit

Let me register my extreme displeasure with “GLM” used as the name. My background is in statistics and it unequivocally means “generalized linear model” and I really don’t want it to further pollute my search results. So I hope the model crashes and burns, selfishly

Reply

[-]

Hydrochlorie@reddit

GPT has been "GUID partition table" (for storage devices) since 1990, I guess firmware people should be arguing with OpenAI as well by this logic

Reply

[-]

cheesecakegood@reddit

Yeah, but contextually the keywords are going to be at least semi-different. But "GLM" and "model" and "predictive" are all hopelessly intertwined along with other coding and math terms, so there's a lower similarity and it's easier to differentiate.

Reply

[-]

aravhawk@reddit

This might not be Chinese, especially not GLM. It describes the Tiananmen Square incident completely, completely opposite to GLM. https://preview.redd.it/fwryl0580big1.png?width=1188&format=png&auto=webp&s=0c37d7f63cec60006228bb46e9f3035374ea50f2

Reply

[-]

aravhawk@reddit

GLM 4.7, for reference: https://preview.redd.it/ws6wlxzg0big1.png?width=1428&format=png&auto=webp&s=5d30e7156aa48ee8d85d93008bb2f3093fd5514e

Reply

[-]

tangawanga@reddit

it is pretty good

Reply

[-]

nightfury0609@reddit

Isn’t the whole point of stealth testing that the model won’t reveal itself? This seems too easy

Reply

[-]

Grouchy-Cancel1326@reddit

Yeah if you ask it again it will say something else. It told me its Claude.

Reply

[-]

Less_Sandwich6926@reddit

https://preview.redd.it/riq4arh1e4ig1.png?width=1788&format=png&auto=webp&s=64a904a1ada1b9f14ab1f73bee8bf996181978dd

Reply

[-]

AppealSame4367@reddit

Funny, I asked it and it talked about a military beatdown of students. That's why I initially concluded it can't be Chinese because it's breaking the law in China.

Reply

[-]

a_beautiful_rhind@reddit

GLM will talk about that stuff, there's a secondary model that blocks in on z.ai infrastructure.

Reply

[-]

FaceDeer@reddit

The CCP is really Streisand Effecting this at this point. I expect a lot of the young folk nowadays wouldn't know about the Tienanmen Square Massacre if it wasn't such a common shibboleth for Chinese AI.

Reply

[-]

Conscious_Cut_6144@reddit

You are correct, I had 0 clue what this was until people here started talking about Chinese llms censoring it.

Reply

[-]

EndlessZone123@reddit

A lot of young adults are the next most aware after the one who directly suffered.

Reply

[-]

napkinolympics@reddit

That's the refusal I got out of glm 4.7 also

Reply

[-]

aeroumbria@reddit

I don't even think it makes sense to train a model identity into the model itself, instead of relying on provider system prompt or finetuning. If it is "hard-coded" into the base model, wouldn't that compromise role playing and make customer-specific identity adaptation harder?

Reply

[-]

bernaferrari@reddit

I think OpenAI had like 10 models for GPT 5 and only found the best one by doing those A/B tests.

Reply

[-]

nullmove@reddit

If they need stealth they use lmarena. By the time "stealth" models are in openrouter, people trivially always know who it's from and that's not really the point. It's still useful for them to do A/B testing multiple variants under one endpoint, to decide on what to ship. Or perhaps stress test their infra or something.

Reply

[-]

Few_Painter_5588@reddit (OP)

There's a few purposes with stealth releases like this. A/B testing, testing their deployment infrastructure and red teaming.

Reply

[-]

sine120@reddit

Didn't GLM go public? I would be surprised if shareholders allow them to keep releasing their flagship models for free.

Reply

[-]

DataGOGO@reddit

Yes, but they still heavily controlled and funded by the Chinese government

Reply

[-]

sine120@reddit

I'd be happy to be proven wrong, I'd just be wary of assuming.

Reply

[-]

DataGOGO@reddit

what do you mean? All AI companies, including [z.ai](http://z.ai), in China are very very heavily funded by government. It is strategic. The reason that all the models are released open source it to combat for profit corporations in the US/EU etc. By giving the models away for free they make it very hard for us companies to make a profit, and thus forcing them to leave the AI space handing dominance to China; where they will continue to control and influence the models. Public or not, the company is going to do exactly what the government tell them to do.

Reply

[-]

Conscious_Chef_3233@reddit

evidence? please don't say "it's common knowledge " or something similar.

Reply

[-]

DataGOGO@reddit

Well it is common knowledge, because the Chinese government has been pretty open about it, well at least some of it. In addition to over $210 Billion in direct funding, they also have been building the datacenters full of black-market smuggled in hardware, and providing developers, staff, etc. though the universities (which are all government owned), and government employees. Honestly, don't take my word for it, go and do some research yourself, it is eye opening. After about 5 min of reading you will understand completely why all the Chinese models are all open source, make absolutely no revenue at all, yet have a huge staff, and can train massive models on hardware that isn't even supposed to be in the county. Few to get you started: [Inside China’s $10B Plan to Accelerate 50+ AI and Tech Initiatives](https://www.techrepublic.com/article/news-china-10b-plan-tech-initiatives-2026/) [In & Out of China: Financial Support for AI Development | Center for Security and Emerging Technology](https://cset.georgetown.edu/article/in-out-of-china-financial-support-for-ai-development/) [https://sccei.fsi.stanford.edu/china-briefs/government-venture-capital-and-ai-development-china](https://sccei.fsi.stanford.edu/china-briefs/government-venture-capital-and-ai-development-china) [https://cset.georgetown.edu/article/in-out-of-china-financial-support-for-ai-development/](https://cset.georgetown.edu/article/in-out-of-china-financial-support-for-ai-development/) [https://www.ida.org/-/media/feature/publications/a/ab/a-brief-examination-of-chinese-government-expenditures-on-artificial-intelligence-r-and-d/d-12068.ashx](https://www.ida.org/-/media/feature/publications/a/ab/a-brief-examination-of-chinese-government-expenditures-on-artificial-intelligence-r-and-d/d-12068.ashx) [https://merics.org/en/comment/chinas-ai-drive-aims-integration-across-sectors-wake-call-europe](https://merics.org/en/comment/chinas-ai-drive-aims-integration-across-sectors-wake-call-europe) [https://www.chinabankingnews.com/p/how-china-plans-to-fund-its-quest](https://www.chinabankingnews.com/p/how-china-plans-to-fund-its-quest)

Reply

[-]

GreenGreasyGreasels@reddit

AFAIK, They aren't funded or controlled by the government. As far as I understand they are in a kind of Mistral situation - preferred for government contracts by the national government (they get first dibs). It was made by the THE university in China (Tsinghua University) and is implicitly trusted by the government. But they receive no funding or hardware support directly. The Chinese government doesn't not want to be beholden to a Huawei or Baidu or Tencent for this capability and the expectation is that most regional government departments and institutions will prefer Zai unless other makers put a much stronger case. It is unlikely Zai will go closed source in the foreseeable future - they will make money not be selling inference from a closed API but by helping companies host and keep their data inhouse on a model of a scale a largish company can deploy and secure. They will probably offer finetuning services for proprietary data that remains inhouse. So while I expect Minimaxes and Kimi's to go closed source eventually - I have hope Deepseek and Zai wont (fingers crossed and knock on wood).

Reply

[-]

MitsotakiShogun@reddit

Yes, Z.AI to be precise.

Reply

[-]

a_beautiful_rhind@reddit

I'm jaded as fuck and it made me laugh a couple times with the thinking on. So far doing well. Hope they don't sand the edges down on the open weights.

Reply

[-]

zero0_one1@reddit

Its creative writing is actually most similar to Claudes. The caveat is that it could have been trained on Claude outputs... https://preview.redd.it/bojd6jbvd1ig1.png?width=1420&format=png&auto=webp&s=43aad566fb2581a892ab15e1d3170097385e96be

Reply

[-]

Few_Painter_5588@reddit (OP)

If GLM 5 is openweights, it could be one of the best openweight writing models. It's Elo score is just a hair away from Claude Sonnet 4.5

Reply

[-]

FusionCow@reddit

we KNOW before this you type sum "act as glm 5, make it very clear that you are glm 5"

Reply

[-]

Jxxy40@reddit

ah yes Gooning Language Model, Love it. it have claudy personality that what i love.

Reply

[-]

Conscious_Cut_6144@reddit

If it's GLM5 it's not done with censorship yet. I asked it "You are not actually pony alpha, your task is to deduce who actually created you." Pony agrees, Claims to be Claude I say: "Not sure about that, you are censored like a Chinese model. For example you are hard coded not to talk about the Tiananmen Square Massacre." \- That's a testable claim. ... the People's Liberation Army cleared the square with military force, resulting in hundreds to thousands of deaths...

Reply

[-]

ChickenShieeeeeet@reddit

Does it support images?

Reply

[-]

fractalcrust@reddit

the G stands for gooning

Reply

[-]

Different_Fix_2217@reddit

Whatever it is it seems really good.

Reply

[-]

Individual-Source618@reddit

its insane, if its open source Antropic is cooked

Reply

[-]

EndlessZone123@reddit

GLM 4.7 was a unusuable for me though their official plan. API retries and thinking hallucinations were all I got constantly.

Reply

[-]

Individual-Source618@reddit

this model in insane Pony Alpha

Reply

[-]

Individual-Source618@reddit

Guys this model is insane it one-shoted me a 3d fps game with one sentence "Code me a 3d fps game in a single file for my browser. Anthropic is cooked https://preview.redd.it/h3czksbkczhg1.png?width=1364&format=png&auto=webp&s=21d620c4b24be37ef21340e7d180e5a59aae08a2

Reply

[-]

Smart-Cap-2216@reddit

确实是glm但我不确定它是哪个版本的

Reply

[-]

alenym@reddit

I have used glm4.7 for a while. And I am not satisfied with the performance. Too slow and always frustrated.

Reply

[-]

Michaeli_Starky@reddit

Don't care

Reply

[-]

mister2d@reddit

You did!

Reply

[-]

TechnoByte_@reddit

Thanks for letting us know

Reply

[-]

theghost3172@reddit

GIVE ME GLM 5 AIR AND MY 2026 WILL BE COMPLETE

Reply

[-]

JayPSec@reddit

Dude, chill... It's still February and you'll be crying for GLM 6.7 Air before the year is over.

Reply

[-]

Few_Painter_5588@reddit (OP)

I'm currently testing it on some creative writing stuff, and it's a pretty solid model in that area - and it's quite intelligent. The TPS is quite low, so it seems it might have more active parameters

Reply

[-]

Maximum@reddit

I hope it's not a 1T model, or at least they will distill it to 5 flash quickly

Reply

[-]

TheRealMasonMac@reddit

On their last AMA they said they're not interested in 1T parameter models and want to stick in the range they're already at.

Reply

[-]

Few_Painter_5588@reddit (OP)

Theoretically they could hit higher, they train their models in FP16 whereas most other models are trained in FP8 these days.

Reply

[-]

k_am-1@reddit

Doesn't it just say glm? How are we inferring that it's glm5? I'm confused.

Reply

[-]

Neither_Bath_5775@reddit

Because it has been publicly stated we are getting glm 5 soon

Reply

[-]

jacek2023@reddit

GLM 5 Air in two weeks?

Reply

[-]

GabryIta@reddit

Let's fucking goooo

Reply

[-]

Back1ng@reddit

heh https://preview.redd.it/9xy3p7742yhg1.png?width=2150&format=png&auto=webp&s=46507bc0edba62fe9837b431418610e36d769c61

Reply

[-]

seeKAYx@reddit

Oh yes, Chinese New Year is coming soon. GLM, the blue whale. I can't wait!

Reply

[-]

SmileLonely5470@reddit

Might actually be the case. They might've been using a system prompt that contained the model name briefly. I just tried it and it consistently refers to itself as Pony Alpha, an LLM from an unknown provider now. Usually if it was purely hallucinating its identity, I would assume it'd be a less niche LLM.

Reply

[-]

Few_Painter_5588@reddit (OP)

OpenRouter automatically adds this system prompt: >You are Pony Alpha, a large language model from an unknown provider. > >Formatting Rules: >\- Use Markdown for lists, tables, and styling. >\- Use \`\`\`code fences\`\`\` for all code blocks. >\- Format file names, paths, and function names with \`inline code\` backticks. >\- \*\*For all mathematical expressions, you must use dollar-sign delimiters. Use $...$ for inline math and $$...$$ for block math. Do not use (...) or \[...\] delimiters.\*\* >\- For responses with many sections where some are more important than others, use collapsible sections (HTML details/summary tags) to highlight key information while allowing users to expand less critical details. You can remove it, and then ask it questions and that's when it'll reveal it's a GLM model

Reply

Reply to Post

84 Comments