This is their hosted-only model, not an open weight model like they’ve become known for. They got a lot of good publicity for their open weight model releases, which was the goal. The hard part is pivoting from an open weight provider to being considered as a competitor to Claude and ChatGPT. Initial reactions are mostly anger from everyone who didn’t realize that the play along was to give away the smaller models as advertising, not because they were feeling generous.
Comparing to Opus 4.5 instead of the current 4.6 and other last-gen models is clearly an attempt to deceive, which isn’t winning them any points either.
I think there is a moderately large market for models like this that aren’t quite SOTA level but can be served up much cheaper. I don’t know how successful they’ll be in the race to the bottom in this market niche, though. Most users of cheap API tokens are not loyal to any brand and will change providers overnight each time someone releases a slightly better model.
> not an open weight model like they’ve become known for.
Right, they state that they'll release "smaller" variants openly at some point, with few details as to what that means. Will there be a ~300B variant as with Qwen 3.5? The blog post doesn't say.
I wish they had a revenue goal to release openly, that way spending money in them would contribute to better open models in the long run.
This is how I view that the public can fund and eventually get free stuff, just like properly organized private highways end up with the state/society owning a new highway after the private entity that built it got the profits they required to make the project possible.
As a publicity stunt, releasing a 300B open model is pretty smart. You can talk about its strong performance and it being “open” and “available,” but it’s so large that most people can’t use it themselves and might try out the cloud-based offering.
Well, you didn’t post the specs on your rig. I think it’s probably more correct to say that you run it on very beefy but readily available hardware. My point was not that nobody could run a 300B model, but rather that a 300B model is not going to be runnable by a majority of people. Sure, anyone who wants to run that model and has the money to purchase the hardware can do it. But the hardware is going to be pricey and most people don’t already have it unless they were trying to run large models before this. My overarching point is that most people with average laptop specs purchased over the last 3 to 5 years are going to have to consume this from the cloud. Which is great for Qwen.
I just have a 3090 and 64gb ram. Yes this is more than most people have, but calling it a "publicity stunt" is just so uncharitably weird of a characterization.
There's smaller models all the way down too.
Like this should be _exactly_ what we want companies to release.
I apologize. I didn’t mean to suggest a “publicity stunt” was a negative. Perhaps I should have said that it was a great marketing strategy. My point was, they can cite all the metrics associated with a frontier model and yet to actually get those metrics most users will have to purchase cloud-based services. That all. And sure, some people will definitely be able to run the model and benefit from it. As you say, this is what we want.
Yeah it is apparently some kind of marketing strategy I guess. Tbh I can't imagine they're getting enough out of it for it to make sense for them. Personally, I'm not looking the gift horse in the mouth too closely, I'm just happy that the current insane rush to make better models means we get some decent "open" ones to play with.
The large models are actually MoE these days so they're usable on ordinary hardware with weights streaming from SSD, just very slow. You're nonethess right that it makes the cloud-based offering more popular, since you can use that for convenience after testing a few inferences locally.
“Usable but slow” is how you could run regardless of MoE or not, the model architecture has nothing to do with it. MoE might run N times faster than non MoE where N is the number of experts but that’s it.
I'm not interested in adopting an inferior closed source weight from a geopolitical rival. The open source weights argument was the one thing China had going and that I was seriously cheering them on for. They could have been our saviors and disrupted the US tech giants - and if it was open, I'd have welcomed it.
Now they show their true colors. They want to train models on our engineering to replace us, while simultaneously giving nothing back? No thanks. I'd rather fund the shitty US hyperscalers. At least that leads to jobs here.
If there's a company willing develop and foster large scale weights in the open, I'll adopt their tooling 100%. It doesn't matter if they're a year behind. Just do it open and build an entire ecosystem on top of it.
The re-AOLization of the internet into thin clients is bullshit, and all it takes is one player to buck the rules to topple the whole house of cards.
> I'm not interested in adopting an inferior closed source weight from a geopolitical rival. The open source weights argument was the one thing China had going and that I was seriously cheering them on for. They could have been our saviors and disrupted the US tech giants - and if it was open, I'd have welcomed it.
Qwen is not the only Chinese lab, and the others have shown no change in their commitment to open source. Allegedly Qwen hasn't either if their recent statements are to be believed. They're just hoping to capture market share with *-claw customers before releasing an open weights version. We'll have to wait and see how before they decide to release that.
> the others have shown no change in their commitment to open source
I wouldn't call this totally accurate, especially as of late. What's closer to the truth however is that there's lots of second-rate players in China doing open models, that will be getting a lot more attention from local AI proponents if the big names seriously slow down their AI releases. The local AI scene as a whole is quite healthy.
> I wouldn't call this totally accurate, especially as of late.
What exactly has changed? Alibaba just released a bunch of new models and have said 3.6 weights are coming soon. The others labs have shown no signs of slowing down their releases either. Whatever you're referring to is news to me and likely most others.
Whereas I as a Canadian am absolutely eager to see a serious competitor from a rival to the US because sending money south to Anthropic and OpenAI who think it's ok to spy on (or worse) their non-American customers, and are headquartered in a country that is trying to crush my country's economy, interfere in our domestic politics, and put us out of work and making threats on political allies.
I'd prefer them to be open weight, but I'd love to sub a decent competitive coding plan from a European or Chinese provider. Right now they're not quite there. If closing it and charging for it brings them closer to competitive, that's ok.
If the US tech and AI industry long term wants customers and a broad market outside of their own domestic base, they need to reconsider who they are bending the knee to, and how they are defining their policies in relation to the Trump administration.
China (meaning the Chinese government specifically, not the people of course) is widely considered to be a low-key geopolitical rival to the developed West in general including Canada and Europe, not just the U.S. I don't exactly like this and would certainly prefer that this wasn't the case, but we can't exactly ignore the facts. This matters when we choose whom to rely on for things like certain hosted third-party services, including AI inference. GP's stance actually makes a lot of sense from this POV, even though it's just as true that many Chinese folks are doing wonderful work on open-weight local AI.
China has never threatened war against my country; America has. Between the two, it’s clearly safer to lean towards the Chinese options if EU ones aren’t available.
"the developed west" is not really a thing. its not a alliance like the eu just a description for some countries. a lot of them are split over china and its a major political issue in places like poland or germany. the only country where all parties treat china like the enemy is the united states, and thats just because theres only two major ones and both listen to corporations instead of their voters. the eu as a organization is a rival to china (not enemy) with all the special duties and import restrictions but thats just economic self interest and not every member is on board. when you ask the average person i think like 80 percent either dont care or think relations should be more friendly. if you ask people under 25 its basically everyone.
I hate to be the one to tell you this but things changed rather quickly when they elected a clown dictator, and now the US is widely considered to be a low-key geopolitical rival to the 'developed West in general' (blergh) including Canada and Europe ...
Seriously, even if you manage to elect someone capable ever again, the US can't be trusted to not elect someone worse than a toddler again in 4 years. In fact, you can't even be trusted to not elect someone even worse than your current dictator (if you thought it couldn't get worse, Trump isn't the bottom of the barrel by far).
I've been using z.ai and codex latest models since last September.
Each release has been an improvement.
codex handles longer sessions but the quality seems to decline and it tends to over engineer and lose focus. It will happily add slop on top of slop...which may pass immediate tests of "code works" but doesn't pass my criteria of "code as craft"
I'm using z.ai GLM with opencode. It's obvious when GLM loses its mind when the session gets too long.
I've been using AI to support programming for around 3 years now. The models have gotten amazing. However, unless there is a significant breakthrough I have determined that it's best for me to focus on short sessions.
I a) organize my work, b) improve my AGENTS.md, ensure source has appropriate comments to guide the models to the patterns and separation of concerns c) use shorter sessions d) review and test without AI. This approach means I still own my code. The AI is just an assistant.
With this approach GLM-5.1 is an excellent model. I never run out of token allotment on z.ai or codex plans. At this point, I only keep my OpenAI subscription as the ChatGPT desktop app is excellent at long web research tasks and I get codex with it.
You're giving up the rest of your country to a geopolitical rival from a separate region, in a separate hemisphere with smiling expansionist goals, even allowing armed Chinese security to protect Chinese installations in country. So why not give the rest of your country to China.
It will help them get a good flank on the USA such that even when that temporarily embarrassed country gets a leader you, and the rest of the world do like, it will be too late to do anything.
A perfect definition of cutting off your nose to spite your face laid bare for all to see.
"Temporarily embarassed" doesn't even begin to describe what's happening down there.
We have an American neighbour actively funding and amplifying a formerly extremely fringe separatist movement in Alberta -- shades of the Donbas, North American edition --and a US "ambassador" who has the behaviour of a 4chan troll.
The bridge has been blown up. Americans might think they are a midterm election away from salvation, but we're on the whole not so naive.
No, a rational decision based on a crazy man in the US. The US needs to learn, that if it threatens its traditional allies, they go to work with china, the main competitor of the US. If the US wants it allies back, the tariffs have to go, and the childlike rhetoric and threats as well. If not, china _deserves_ the business of the US former allies.
Right, and we're not just watching the behaviour of the US administration, we're watching the behaviour of the electorate / populace. At the polling booths but also in online comment sections, as tourists, consumers, etc.
And mostly not liking what we see. Encouraged by the No Kings protests, but unless that boils over into a hegemonic and stronger opposition, it still seems like there's a 40% population there that can't deal rationally with the world inside their own border, let alone outside.
Also... When Biden took over after Trump's first term most of the protectionist policies stayed and foreign policy didn't really budge (outside of support for Ukraine). I expect similar if (big if) the Democrats regain executive power.
The US under Trump is politically and strategically almost identical to China, and can be trusted about the same.
And then, compared to China, the US acts overtly hostile: threatening us with war, starting a war in order to collapse energy supplies outside of the US.
Opportunistic beyond even China, much more hostile.
Will the US even be a democracy in two years? Is it now?
Nah man, balancing between China and the US is the only thing a smaller country can do in order not to be crushed
> I'm not interested in adopting an inferior closed weights model from a geopolitical rival.
That's a very reasonable stance. It doesn't change the fact that we do have plenty of local models (up to and including Qwen 3.5) that are still quite useful.
looking at the other replies I'm not seeing what I consider the most important rebuttal to this argument: there is no real "adopting" in this hybrid open/close space right now, the lock-in is minimal and as much as different corporations are trying to create a lock-in effect by closing down their tools and interfaces, they are not really succeeding
I can constantly jump from one provider to another, and to my local servers which are already able to run very useful models at reasonable hardware cost, and I intend to continue doing that for the foreseeable future
the one thing I'm not going to do is tying my tooling to one provider or another or getting overtly used to the specifics of a model outside of my control
more than the weights or the training, which of course are very important, the real battle right now is for establishing some dependency mechanism so that your users won't just flee en-masse as soon as you inevitably try to abuse your market dominance and lock-in mechanisms, as is customary in everything computing these days - note that i don't explicitly talk about raking up prices, that is just one of the most difficult methods as people are very sensitive to that, when you can sneakily sell user data, get government contracts from never-disclosed conditions, or even just incorporate your intelligence to ad networks in one way or another
z.ai models are open weights. GLM-5.1 is very close to Opus with obvious exception of session length.
Only academic models will be true open source as companies can't legally afford to disclose learning inputs.
In regards to "They want to train models on our engineering to replace us". Some software engineers in China can run circles around some of the best teams in Silicon Valley. Days of U.S. hegemony are over. I recommend you make peace and make friends.
Interesting! What is your reasoning behind that? I just learned there where closed models from the team before this so that shouldn’t have been a surprise for the employees? Or do you think the internal communication was: we will release better open models the the existing closed ones to push everything forward and now when they are getting competitive they are becoming proprietary?
I feel like this is true. I don't mind being a blip behind the bleeding edge if I don't have to change my tooling every month. But the second my current provider tries to screw me over, I'll still jump ship
The business model, howver, is lobster in a bucket. Any model that starts gaining as a private model will have competitors to release comparable open models because those locked in customers will not swotch unless you demo the capabilities.
So expect every now and then a open model burp from the trailing frontiers. Afterall, its all sunk cost so once you have it and no customers, theres zero reason not to spike your competition and try again or exit.
I use different models in production and model's "personality" as in tendency to not go off script, not consume gazillions of tokens recursively, follow instructions etc, are more relevant than "brute" power which is okayish as a metric for agentic coding on generous token plans.
Chinese models are very competitive in that regard, you'll often look at 70-90% price reduction at the same quality.
I've found Opus 4.6 to be smarter than 4.5, at least in some ways. There's a bug I'd been trying to solve for a decade (and so had other humans) and I've been giving it to each model to try and solve, including in interactive sessions. Each model got closer, but none of them actually solved it, until Opus 4.6 got it on the first go (I probably used Ultrathink). This was before the 1M context was available.
I'd agree that 4.6 and 4.5 are different, but I don't think it's correct that 4.6 is just reduced and benchmaxxed. It genuinely solved problems for me that no other model has been able to.
I think I'd like to have seen the 4.6 benchmarks also included against Qwen.
I’m starting to wonder where the most is for any of these models.
Sure they are not cheap to train. But if open weight models continue to be trained and continue to become available on cheaper hardware, how do dedicated AI companies protect their margins?
OpenAI found the answer: artificially curtail the supply of DRAM wafers (by buying 40% of world's supply without necessarily having a feasible plan to make use of all of them], to prevent consumers from getting access to gpus with more and more memory, which could allow them to get dangerously close to the state of art while running local AI
Rather than an increase in VRAM of consumer gpus, we are seeing a decrease, which is pretty optimal for OpenAI
Opus was released in Feb 2026. Even though it feels like a long 2 months has passed, its' not really clear that they were developing this as a competitor to that product.
There's nothing really strange about not competing directly with the best, but rather showing whom you are as good as.
I don’t know why anyone would do the mental backflips to defend this.
They posted charts with logos for Claude and others. You had to read the fine details to realize they weren’t comparing to the latest offerings from those companies. They were counting on you not noticing.
There’s zero reason to compare to old models unless you’re trying to mislead.
> Initial reactions are mostly anger from everyone who didn’t realize that the play along was to give away the smaller models as advertising, not because they were feeling generous.
The naivety around this has been staggering quite frankly. All of a sudden, people thinking that meta etc are releasing free models because they believe in open access and distribution of knowledge. No, they just suck comparatively. There is nothing to sell. Using it to recruit and generate attention is the best play for them.
I thought Qwen was releasing open-weight because China can't compete with America (because of people's privacy concerns), so the only thing they could do is salt the ground economically with open models, and make sure everybody loses.
Qwen is actually a pretty strong player in the Chinese market. There is an implied "salt the ground" play but it's mostly from hardware makers, who are trying to keep the big AI players honest and also stand to gain if local inference becomes popular.
For a brief moment there were a lot of comments about how Chinese tech companies are our saviors in the age of AI because they were releasing their models. It was an edgy contrarian take that was getting a lot of traction, mostly from commenters who were unfamiliar with Alibaba and thought it was the anti-Big-tech
I'm not frustrated or disappointed, we have lots of models from Qwen already. We haven't really lost anything. And plenty of players only release "smaller" models anyway, so it's hardly unprecedented.
OP didn't say about confusing Opus with Qwen but rather people being confused about Qwen3.6-Plus not being available as an "open weight" model available for self hosting.
No. Right now I'm upset that Google has removed (or at least is in the process of removing) the Gemini 2.0 flash model. We use it for some pretty basic functionality because it's cheap and fast and honestly good enough for what we use it for in that part of our app. We're being forced to "upgrade" to models that are at least 2.5 times as expensive, are slower and, while I'm sure they're better for complex tasks, don't do measurably better than 2.0 flash for what we need. Yay. We've stuck with the GCP/Gemini ecosystem up until now, but this is kind of forcing us to consider other LLM providers.
this is one of the reasons im hearing more and more people are using open/locally hosted models. particularly so we dont have to waste time to entirely redo everything when inevitably a company decides to pull the rug out from under us and change or remove something integral to our flow, which over the years we've seen countless times, and seems to be getting more and more common.
products entirely disappearing or significantly changing will be more and more common in the llm arena as things move forward towards companies shutting down, bubbles deflating, brand priorities drastically reshifting, etc...
i think, we're at or at least close to a time to really put some thought into which pieces of your flow could be done entirely with an open/local model and be honest with ourselves on which pieces of our flow truly needs sota or closed models that may entirely disappear or change. in the long run, putting a little bit of thought into this now will save a lot of headache later.
Yeah. Back when Gemma2 came out we benchmarked it and were looking at open models. For our use case though, while the tasks are pretty simple, we do need a pretty large context window and Gemini had a big lead there over the open models for quite a while. I'll probably be evaluating the current batch of open models in the near future though.
Thanks. Yeah, for now we're moving to 3.1 flash lite as that's the new cheapest at $.25/1M and is also still "good enough". 2.5 flash is more expensive at $.30/1M (looks like Deep Infra charges the same as GCP/VertexAI for it). I might check them out for Gemma though. We benchmarked Gemma2 when that came out and it wasn't remotely usable for us largely because the context window was way too small. It looks like 3 or 4 might be worth evaluating though.
OpenRouter usage is likely skewed towards LLMs that are more niche and/or self-hostable by solid hardware that's available, but most consumers don't have on hand. I can imagine Anthropic and OpenAI LLMs often get called directly from their APIs instead.
At least from my experience and friends of mine, we use OpenRouter for cases where we want to use smaller LLMs like Qwen, but when I've used ChatGPT and Claude, I use those APIs directly.
0.1% of OpenRouter is around 400 billion tokens per month or around $400k per month at a cost of $1 per 1 million input tokens, not counting output.
I think it's pretty disingenuous to call your SaaS little when it is projected to spend at least 5 million USD just on tokens and this is a low end estimate.
Their homepage says 30T tokens monthly, so 0.1% would be 30 billion.
And I pay way less than $1 per input token, especially when caching is taken into account.
EDIT: they updated it in the last day or two, now it says 70T, so I’m a little below 0.1% now. But seriously, the point stands, 70T tokens a month just isn’t that much in the global scheme. The big labs are pushing quadrillions each.
> There isn't, pretty much everyone wants the best of the best.
For direct user interaction or coding problems, perhaps. But as API calls get cheaper, it becomes more realistic to use them for completely automated workflows against data-sets, or as sub-agents called from expensive SOTA models.
For example, in Claude, using Opus as an orchestrator to call Sonnet sub-agents, is a popular usage "hack." That only gets more powerful, as the Sonnet equivalent model gets cheaper. Now you can spawn entire teams of small specialized sub-agents with small context windows but limited scope.
I did create my own MCP with custom agents that combine several tools into a single one. For example, all WebSearch, WebFetch, Context7 exposed as a single "web research" tool, backed by the cheapest model that passes evaluation. The same for a codebase research
Use it with both Claude and Opencode saves a lot of time and tokens.
Smart approach combining tools behind a single interface. The "cheapest model that passes evaluation" pattern is underrated.
One data point that might help with the research tool: not all APIs work equally well when called by agents. We scored 387 APIs on agent-friendliness and 54% fail the bar. The main gaps are no CLI tool (66%) and no machine-readable pricing (72%). If your research tool is helping agents pick APIs to integrate, the scores at clirank.dev/api/apis?sort=score could save the expensive model from wasting tokens on APIs that'll fail headless.
> But as API calls get cheaper, it becomes more realistic to use them for completely automated workflows against data-sets
Seems like a huge waste of money and electricity for processes that can be implemented as a traditional deterministic program. One would hope that tools would identify recurrent jobs that can be turned into simple scripts.
For example: "Here our dataset that contains customer feedback comment fields; look through them, draw out themes, associations, and look for trends." Solving that with a deterministic program isn't a trivial problem, and it is likely cheaper solved via LLM.
It makes sense if the dataset is so large that LLM cost is a prohibitive factor. Otherwise a frontier LLM has the advantage of producing a better result.
Not all tasks require models like opus. If they do not, then it is more efficient to use cheaper and faster models. For most of my tasks now I use the big kimi/qwen/glm models because they are cheap and good enough, if not even the smaller locals ones.
I would say that for a significant part of the current market open-source models are good enough to fill a part of it.
maybe there isnt, but as understanding grows people will understand that having an orchestration agent delegate simple work to lesser agents is significant not only for cost savings, but also for preserving context window space.
That isn't true. In a Codex or Claude Code instance, sure... but those are not the main users of APIs. If you are using LLMs in a service for customers, costs matter.
The market for API tokens is bigger than people like you and I (who also want the best) using then for code.
There are a lot of data science problems that benefit from running the dataset through an LLM, which becomes bottlenecked on per-token costs. For these you take a sample subset and run it against multiple providers and then do a cost versus accuracy tradeoff.
The market for API tokens is not just people using OpenCode and similar tools.
Nope. I get very good results from GLM 5 and 5.1. I’m not working on anything so complex and groundbreaking that I need the best.
Coding is a rung on the ladder of model capability. Frontier models will grow to take on more capabilities, while smaller more focused models start becoming the economical choice for coding
Not really. It depends on the usecase. For private stuff I'm very happy to take what was SOTA a year or 2 ago if I can have it all running in my home and don't have to share any of my data with some sleazy big tech cloud.
The price is a concern too of course. But privacy is a bigger one for me. I absolutely don't trust any of their promises not to use data for training purposes.
Comparing to Opus 4.5 instead of the current 4.6 and other last-gen models is clearly an attempt to deceive, which isn’t winning them any points either.
I think there is a moderately large market for models like this that aren’t quite SOTA level but can be served up much cheaper. I don’t know how successful they’ll be in the race to the bottom in this market niche, though. Most users of cheap API tokens are not loyal to any brand and will change providers overnight each time someone releases a slightly better model.