ZAI might stop open-weighting their models?

Posted by TheRealMasonMac@reddit | LocalLLaMA | View on Reddit | 60 comments

Ever since the company went public, they’ve been making a lot of changes that clearly seem to be prioritizing profit without regard to their customers. For example, with their coding plans: \- They promised/advertised that the Lite coding plan would always have the most up-to-date models, but then silently omitted GLM-5 from the plan. \- They’ve raised prices twice without any accompanying announcement or acknowledgement. Their latest price increase plops them in the range of Anthropic/OpenAI’s subscription plans. \- They abruptly and silently backtracked on a policy that allowed subscribers to use their subscription for coding tools of their choice, and are now even trying to gaslight you into believing their own documentation doesn’t say you can (it still says so)! Furthermore, they stopped releasing their base models. GLM-4.7-Flash and GLM-5 do not have an accompanying base model release. Looking at this trajectory, I fully expect that they will stop releasing their top models as open-weight soon enough (if not with their next launch). Whatever new management they onboarded seems more than willing to destroy any goodwill towards the company. And it clearly doesn’t have to be this way! Looking at MoonshotAI and MiniMax, who are both in similar positions, they haven’t abruptly screwed over their customers and—get this—actually communicate. It’s fair-game to make business decisions to turn a profit, but there’s no reason it has to be done like a douchebag.

Reply to Post

60 Comments

[-]

FullOf_Bad_Ideas@reddit

Minimax and Zhipu both IPOed in January 2026. Check their stock prices. They're under big pressure to deliver revenues that will support them. Meaning that pe using their open weight models locally or on self-serve basis will get screwed.

[-]

Alternative_You3585@reddit

Compute is not free. I doubt they will stop making open weights, but they certainly will not host it themselves for free much

[-]

putrasherni@reddit

what benefit do they get by making their models open weight ?

[-]

Alternative_You3585@reddit

China is following communistic principles and it's working great, it's more about research than modelsm their advances complimentary each other which allows them to develop quite quickly

[-]

RepulsiveRaisin7@reddit

As someone with communist-adjacent beliefs, literally none of this is communism.

[-]

FastHotEmu@reddit

Agreed... China is more "State Capitalism" than anything.

[-]

Guinness@reddit

Damaging the United States. The more economic harm they can do to US tech sectors the better (from their point of view). Think about it. The US dominated tech for how long? LLMs are the next big thing. And it was looking like the US was going to dominate it yet again. And then DeepSeek dropped. It’s hugely beneficial to the CCCP.

[-]

FastHotEmu@reddit

This. Flooding the market with cheap alternatives is exactly what they are doing.

[-]

05032-MendicantBias@reddit

An army of free labour fine tune, port and improves their models.

[-]

Daemontatox@reddit

Publicity, sponsors, investors....etc

[-]

RuthlessCriticismAll@reddit

> sponsors, investors show your work

[-]

PinkySwearNotABot@reddit

"trust me bro"

[-]

Daemontatox@reddit

Well if i was a company like zai or qwen for example , i can just create the model , run the benchmarks and show how good it is aswell the report paper thats published as my contribution to the open-source community which alot of people dont pay attention to even though its really a big deal. I can then distill this model into a really small model in comparison lets say 700->7/30/100B and then open-source it and say i am giving to the community. I am then able to keep the best of two worlds , i kept my best for investors and big corps that want to maybe collab aswell , i helped the community by giving them toy models and a paper which indirectly helps me on the long run by allowing more minds to tackle an issue by maybe finding a new way to finetune or new distillation or quantization which i can then use for my next project and rinse and repeat. If i remember correctly deepseek did this very well with their optimized pipelines for data parallelism and hermes were the first to train / finetune a model hsing somekind of RL (not too sure pls dont kill me) , so its more of i scratch your back and you scratch mine.

[-]

Dabalam@reddit

There is an argument about their ability to capture market share and notoriety. Publicity is good and I imagine the training models being a gateway drug to the larger closed models might be a strategy that works (assuming Qwen are right about this strategy).

[-]

arcanemachined@reddit

Also, Moonshot recently allowed Cursor to finetune Kimi K2.5 to make their proprietary model. So that could presumably be a business case as well: Provide a base model, and sell the ability to fine tune on top of it.

[-]

MikeFromTheVineyard@reddit

Most of the world wouldn’t try a Chinese closed model over a superior closed US model. Opening them gives them plenty of attention and attracts users. Once they build a reputation, that incentive diminishes

[-]

ProfessionalSpend589@reddit

My upvote.

[-]

nivvis@reddit

hosting it for free is not for generosity .. where do you think all of these traces they are training off of are coming from? imo there's not a lot of downside in releasing openly. real models are continuously RLHF & RLVR'ed and take a lot of setup behind the scenes to serve for real -- to say it's not totally trivial to actually compete with them using their model. And tbf they don't have to release every checkpoint etc to get the benefits of open source (and most likely aren't .. just releasing their larger versions etc). Just playing devil's advocate. Hard to say what their real endgame is .. probably whatever gets them the strongest models so they just keep some upper hand.

[-]

Thomas-Lore@reddit

This indicates they are heavily compute starved, so why would they stop releasing local models? Having other providers saves them cimpute while keeping the model popular. They may go the minimax way though of having agreements with the providers, and only allowing home use without it. IMHO.

[-]

SexyAlienHotTubWater@reddit

What's the profit model spending loads of money to train GLM and then releasing it for free?

[-]

TheRealMasonMac@reddit (OP)

I thought about that, but then it makes no sense for them to still explicitly allow OpenClaw.

[-]

rpkarma@reddit

How do they enforce it? Still working for me, you just need to use the Anthropic compatible endpoint

[-]

TheRealMasonMac@reddit (OP)

I don’t know. But this is their automated email response about the policy (according to someone else): ``` Thanks for reaching Z.ai Maas IT support. Our normal working hours are Monday - Friday 9:00 - 18:00 UTC+8. Below are the details for your issue: Issue Definition: Error 1313 - Temporary restriction triggered by system protection policy Current Situation: We have detected usage behavior that may not align with normal individual usage patterns. As a result, certain subscription functions have been temporarily restricted. This is an automatic protection mechanism designed to ensure platform stability and fair usage across all users. In most cases, this type of restriction is temporary. You may review your recent usage behavior and check whether any of the following situations apply: 1. **Using unofficial methods to invoke the Coding plan: Bypassing coding tools and directly calling the https://api.z.ai/api/coding/paas/v4/chat/completions endpoint via curl or other methods is considered a violation.** This endpoint is not documented in our official guidelines and is not the correct way to use the Coding plan. Please configure your tools according to the official documentation. 2. Abnormally high-frequency requests: Sending an extremely large number of requests in a short period will be flagged as a malicious attack, resulting in an account ban. 3. Account sharing: Suspicious activities indicating that multiple users are sharing a single account. 4. Unauthorized reselling: Accounts suspected of selling or transferring Coding plan quotas without authorization. After this, no additional action is required from your side. The system will continue to monitor account activity and will automatically reassess and restore access once normal usage conditions are met. If you have any further questions or requests related to the case, please do not hesitate to contact us. Best regards, Z.ai MaaS Developer Support Working Hours: Mon-Fri 9:00 - 18:00 UTC+8 ```

[-]

rpkarma@reddit

Yeah that’s the OpenAI endpoint. Can’t use that now, it’ll reject you. The anthropic one is what OpenCode etc use, so I connected Pi and my own agent to it: no worries so far

[-]

jacek2023@reddit

It's a good thing because it will be more difficult to fake local users to pretend they use big models locally

[-]

a_beautiful_rhind@reddit

They are hard up for money. Yesterday 4.7 disappeared from their site. Now you only get 5.1, 5V and 5-turbo.

[-]

TurnUpThe4D3D3D3@reddit

They’ll keep open sourcing unless they manage to create a frontier model.

[-]

HornyGooner4402@reddit

Lite plans definitely have access to 5 now and even 5.1 on launch. I don't know why you expected them to let everyone use a new model that's twice the size as if they'd magically have twice the compute. The only model they didn't release was the 5 Turbo and they still allow other coding tools and even OpenClaw, though coding use will be prioritized. Lastly, they still respect the price and plan you bought it for, even my legacy plan still allows ~30M tokens every 5 hours with no weekly limit for just $72 a year. Did you pull all of this info out of your ass?

[-]

SillyLilBear@reddit

Minimax is going same direction

[-]

Ok_Warning2146@reddit

Not open weight any more won't kill them. But the way they treat their paid customers will get them killed by competitors

[-]

openclaw-lover@reddit

What would you do if you were in their shoes?

[-]

Widget2049@reddit

make a smear campaign about it in reddit for other models! just like the one op did, it's a very good template.

[-]

TheRealMasonMac@reddit (OP)

Okay, Mr. 1 post karma 188 comment karma who is definitely not a bot.

[-]

FailedTomato@reddit

They have to prioritise profits to be a sustainable organisation. Nothing mysterious about that. The aggressive open weights releases were important for publicity. I'm confused about their pricing though. Their models are clearly inferior to Claude, even on benchmarks. Why would people pay the same subscription price? Do they expect Claude prices to drive up or usage quotas to go down even further?

[-]

rpkarma@reddit

Claude usage quotas have gone down further already. The Pro sub is nearly useless.

[-]

FailedTomato@reddit

Yes. I guess useless for serious software engineering work. I'm a researcher and I find Claude pro to be still be very useful. Not sure how long that will last tho lol.

[-]

rpkarma@reddit

Yeah I should be clearer: far less useful for software dev. A single prompt/task on Opus blew through my entire quota for that period, it’s insane. No change in the tasks I’m giving it either, they messed with something on their side.

[-]

TheRealMasonMac@reddit (OP)

I am criticizing the fact the made changes silently, which implicitly means they believe they've reached a critical mass where they can make such decisions without announcement (again I'm criticizng the lack of communication, not the decision itself) regardless of how it destroys goodwill. What's stopping them from eventually thinking, "Hey, I actually don't need to open-weight my top models anymore." Enough people seem to be getting the impression I'm criticizing [Z.AI](http://Z.AI) for chasing profits that I am genuinely puzzled. I'm not. Can you explain how you yourself arrived at that interpretation so I can clarify the post?

[-]

FailedTomato@reddit

Your post is fine. I think people are not surprised by companies shifting away from open weights. It's happening across the board. I also agree that these silent changes are damaging to their reputation. And this is the surprising part I think. That they think they can afford to do this, as well as charge frontier model prices.

[-]

mitchins-au@reddit

What do you expect from a company. Apart from Google who can afford to crank out Gemma on the side, money has to come from somewhere. They’re not a charity, don’t be surprised.

[-]

TheRealMasonMac@reddit (OP)

I'm not lol. But a lot of people seem to be under the assumption that Chinese companies release models from the goodness of their hearts and will keep doing so.

[-]

mitchins-au@reddit

Only if they are sponsored to do so

[-]

TheRealMasonMac@reddit (OP)

Yeah, but that incentive seems to be drying up. Alibaba recently shifted towards proprietary models to a greater extent than before. Withholding Qwen3.6-397B is possibly part of that unless they plan to release 3.6 all at once. Src: https://www.ft.com/content/b39da303-3188-447b-8b65-3dd8dad8b59a?syn-25a6b1a6=1t

[-]

mitchins-au@reddit

It’s not like anyone even here is running 397B at home. If they’re making money hosting it, the only people running it would be inference providers under cutting them

[-]

rpkarma@reddit

Speak for yourself lmao.

[-]

ttkciar@reddit

> It’s not like anyone even here is running 397B at home. What? Some of this sub's regulars have been running models larger than that on their homelab hardware since 2024.

[-]

rpkarma@reddit

The entire Chinese scene is rapidly trying to make money now. Alibaba deleted their Lite plan entirely for those of us still on it yesterday too: no more renewals. Existed for two months lol These companies are realising that there’s no profit here unless they jack the prices way up and close up shop.

[-]

SteppenAxolotl@reddit

>prioritizing profit without regard to their customers Did all the inference providers that sell API tokens to their open models give them a cut? It was free inputs to their business and it never crossed their minds what happens if the source of their free inputs went away. Traditional definition of a customer is: One that buys goods or services, as from a store or business.

[-]

TheRealMasonMac@reddit (OP)

So, are you in agreement with me or making an argument against...? I'm genuinely unsure because that doesn't seem to actually push back against anything I said. Customers, here, refers to customers of ZAI. Not to any inference provider. I am criticizing the fact the made changes silently, which implicitly means they believe they've reached a critical mass where they can make such decisions without announcement (again I'm criticizng the lack of communication, not the decision itself) regardless of how it destroys goodwill. What's stopping them from eventually thinking, "Hey, I actually don't need to open-weight my top models anymore."

[-]

Final_Ad_7431@reddit

the models get bigger, mote intense, they get a lot more competitive against the big players in the space, it's very hard for a company to close off, package up and sell like a <= 200b model because aside from the minimax's and stuff it doesn't really compete so you may as well make it open for reputation/good will/free advertising etc (i think qwen has done great at this) but now we're at the point where glm5.1, im guessing kimi k2.6 will be up there, minimax m2.7 seems pretty good for it's size - all of these are starting to compare pretty well against things like opus, they might not match but they're getting close enough while being cheaper that it makes sense to start thinking about closing off and monetizing them more heavily, i don't like it obviously but i can understand it

[-]

AltruisticList6000@reddit

I'd be happy if they end up closing the huge models, at least they'd keep releasing smaller open weight models like 9-22b range dense models or similar like Mistral Small (Mistral released a bigger model recently though for the first time since ages), that way they can monetize but also be kind to us local users which is also a marketing win over fully closing down everything.

[-]

Embarrassed_Adagio28@reddit

I am confused about a few things in your post. I have the z lite coding plan and can use glm 5.1 with opencode without issues? However I went to look at how much it would cost to upgrade and now the lite plan is $18 a month and the middle tier is like $72. My account says I only owe $10 a month so at least my price didn't rise but damn.

[-]

TheRealMasonMac@reddit (OP)

You can't use GLM-5.

[-]

dark-light92@reddit

Ever heard about compute constraints?

[-]

TheRealMasonMac@reddit (OP)

Ever heard of whataboutism?

[-]

dark-light92@reddit

Yes. This post is prime example.

[-]

LittleYouth4954@reddit

I have a Lite plan and access to glm 5.1 and 5v. Why would someone still use glm 5? Also, their prices are still cost-effective considering the quality of the models and the volume of tokens. Finally, glm 5.1 has open weights.

[-]

TheRealMasonMac@reddit (OP)

It's the trajectory they seem to be following.

[-]

LittleYouth4954@reddit

You don't have to support your hypothesis with misinformation

[-]

TheRealMasonMac@reddit (OP)

Are you maybe not a native English speaker? That's not what "misinformation" means. Misinformation means they are false or untrue. Everything I said are objectively true. Hypotheses are supported by facts. Therefore your point doesn't really make sense.