> Cursor Composer 1 was Qwen and this is Kimi. IDE is based on VSCode. The entire company is build on packaging open source and reselling it.
The question is, where's the outrage? Why are there no headlines "USA steals Chinese tech?" "All USA can do is make a cheap copy of Chinese SOTA models".
> So funny to see Twitter go wild saying "a 50 person team just beat Anthropic" blah blah.
Well, if it's an American company, then it's a noble underdog story. When Chinese do it, they are thieves leeching on the US tech investment.
Do you think Chinese LLMs acquired training data legitimately? I think the whole situation is a bit funny, but I don't think the US "started it" to be fair.
> I think they probably acquire it in accordance with Chinese law.
You can easily look up[1] how China struggles with effective enforcement of IP laws.
And specifically for LLMs, Anthropic recently claimed that Chinese models trained on it without permission.[2]
> Who are you quoting with those marks?
Double quote marks have other uses besides direct quotes, such as signaling unusual usage.[3] In this case, talking about countries like they're squabbling kids.
> Started what?
Fishy use of others' IP, packaging others' work without attribution.
> To be fair to whom?
To US companies using Chinese LLMs without attribution.
They said Chinese law, which is not the same as American law, and presumably using IP the way they have is legal there, if indeed they actually did, as allegations of IP theft are just that, allegations, and even if they weren't, all nations in the history of mankind have been "stealing" "intellectual property" since forever, including the US from Britain, literally with the good graces of the fledgling US government [0].
As to what Anthropic said, it's quite specious as this analysis shows [1], ie the amount of "exchanges" is only tantamount to a single day or two of promoting, not nearly enough to actually get good RL training data from. Regardless, it's not as if other American LLM companies obtained training data legitimately, whatever that means in today's world.
> You can easily look up[1] how China struggles with effective enforcement of IP laws.
I didn't see anything in there about Chinese companies violating Chinese law.
Can you so easily look up how American companies struggle with effective enforcement of Chinese IP laws? I think it should be pretty easy to see how American companies struggle with effective enforcement of European IP laws, and I can tell you it is similar.
From here, it is not so clear that the US can even enforce its own laws at the moment.
> signaling unusual usage
Thank you!
> In this case, talking about countries like they're squabbling kids.
> > Started what?
> Fishy use of others' IP, packaging others' work without attribution.
I see. I guess if China is 3000 years old then maybe obviously, because the US is such a young country by comparison.
So you think it is "fair"[1] to violate Chinese Law because there were people in China who violated US law first?
I mean, I (and a ton of others) were pretty outspoken about ollama being a pack of grifters. The thing they are good at is marketing though, so it drowns out other projects in the area.
yup. fully agree. American cry and bitch about Chinese copy and steal their tech then an American company (Cursor) use/steal open source tech from China and everyone is silence.
It's a bit more than that. They have plenty of data to inform any finetunes they make. I don't know how much of a moat it will turn out to be in practice, but it's something. There's a reason every big provider made their own coding harness.
Can anyone enlighten me how having a coding harness when for most customers you say "we won't train on your code" helps you do RL? What's the data that they rely on? Is it the prompts and their responses?
Eg, When a prompt had a bad result and was edited, or had lots of back and forth to correct tool usage that information can be distilled and used to improve models.
And now imagine if you are focused on this for weeks you can likely come up with other ideas to leverage the metadata to improve model performance.
It doesn't matter what your privacy setting is, with any savvy vendor. Your data is used to train by paraphrasing it, and the paraphrasing makes it impossible to prove it was your data (it is stored at rest paraphrased). Of course the paraphrasing stores all the salient information, like your goals and guidance to the bot to the answer, even if it has no PII.
That's an interesting accusation there! You're essentially accusing every "savvy vendor" of large-scale fraud... DOn't suppose you'd have any actual citations or evidence to back that up?
Cursor’s integration is much deeper than just plugging an LLM into VSCode
That said I have a feeling both VSCode and Claude code will catch up to their integration. But neither comes close yet (I say that as someone who mainly uses Claude Code).
You can copy/paste or drag code snippets the chat window and they automatically become context like. (@myFile.cpp:300-310)
You can click any of the generated diffs in the assistant chat window to instantly jump to the code.
Generated code just appears as diffs till you manually approve each snippet or file. (which is fairly easy to do with "jump to next snippet/file" buttons)
These are all features I use constantly as someone who doesn't vibe but wants to just say "pack/unpack this struct into json", "add this new property to the struct, add it to the serialization, and the UI", and other true busywork tasks.
I used Cursor for the second half of last year. If you’re hand-editing code, its autocomplete is super nice, basically like reading your mind.
But it turns out the people who say we’re moving to a world where programming is automated are pretty much right.
I switched to Claude Code about three weeks ago and haven’t looked back. Being CLI-first is just so much more powerful than IDE-first, because tons of work that isn’t just coding happens there. I use the VSCode extension in maybe 10% of my sessions when I want targeted edits.
So having a good autocomplete story like Cursor is either not useful, or anti-useful because it keeps you from getting your hands off the code.
I use both pretty heavily. Cursor has an "Ask" mode that is useful when I don't want it to touch files or ask a non-sequitur. Claude may have an easy way to do this, but I haven't seeked it.
Cursor also has an interesting Debug mode that actively adds specific debug logging logic to your code, runs through several hypotheses in a loop to narrow down the cause, then cleans up the logging. It can be super useful.
Finally, when making peecise changes I can select a function, hit cmd-L and add certain ljnes of code to the context. Hard to do that in Claude. Cursor tends to be much faster for quicker, more precise work in general, and rarely goes "searching through the codebase" for things.
Most importantly, I'm cheap. a
If I leave Cursor on Auto I can use it full time, 8 hours a day, and never go past the $20 monthly charge. Yes, it is probably just using free models but they are quite decent now, quick and great for inline work.
The majority of Ask/Debug mode can be reproduced using skills. For copying code references, if you're using VS Code, you can look at plugins like [1], or even make your own.
Cursor's auto mode is flaky because you don't know which model they're routing you to, and it could be a smaller, worse model.
It's hard to see why paying a middleman for access to models would be cheaper than going directly to the model providers. I was a heavy Cursor user, and I've completely switched to Codex CLI or Claude Code. I don't have to deal with an older, potentially buggier version of VS Code, and I also have the option of not using VS Code at all.
One nice thing about Cursor is its code and documentation embedding. I don't know how much code embedding really helps, but documentation embedding is useful.
Mostly saying "include this line from x file and this block from y file" which keyboard shortcuts. Claude's VSCode plugin only does one selection. Claude Code requires explicitly telling it what to reference.
That plus Cursor's integration into VSCode feels very deep and part of the IDE, including how it indexes file efficiently and links to changed files, opens plans. Using Claude Code's VScode extension loads into a panel like a file which feels like a hack, not a dedicated sidebar. The output doesn't always properly link to files you can click on. Lots of small stuff like that which significantly improves the DX without swapping tabs or loading a terminal.
I also use Code from terminal sometimes but it feels very isolated unless you're vibecoding something new. I also tried others: Zed is only like 50% of the way there (or less). I also tried to use (Neo)Vim again and it's also nowhere close, probably 25% of the UX of Cursor even with experimental plugins/terminal setups.
It just means Kursor is sharing data with Chinese llm which enables them to improve their LLM by training on outputs and input of all data which cursor collects.
We know Composer 2 is Kimi K2.5 from that tweet. Where is the evidence for Composer 1 being based on Qwen?
> So funny to see Twitter go wild saying "a 50 person team just beat Anthropic" blah blah.
In this case, it will be the other way round: Anthropic will see Cursor as a competitor AI lab using open weight models for Composor 2 (actually Kimi K2.5) which was allegedly distilled from Opus 4.6, and would be enough for Anthropic to cut off Cursor from using any of models.
How does this blow that narrative up? A 50 person team likely broke a license to have a product that's competitive on output at a fraction of the costs of one of the most well capitalized companies on the planet. Claude code and anthropic are certainly the darlings of the space today, but to me this just reinforces the idea that their moat is razor thin on the model front, even compared to OSS that can be run on independent hardware.
The application layer play is also suspect to me. In the medium to long term I _want_ tools that'll let me run whatever models I want vs being tied to an expensive, proprietary, and singular provider. For personal work I care about costs, and eventually my employer will care both about costs _and_ enterprise features/governance that a company like Anysphere is extremely well positioned to provide.
More and more, I see the future of the application layer being model agnostic, most enterprises hosting models on their own cloud for data security concerns, and the models being fully commoditized.
Considering how AI companies incestously RL on each other's models, I would not be surprised if any number of behavioral patterns and (claims to be ChatGPT/Claude/Deepseek or whatever) just popped up on new models constantly.
I would also not rule out that since K2 is an 1T model, this is a distill, as I don't think they're serving expensive models just like that, which would not be a licensing violation?.
There's a now-deleted tweet from a Kimi dev claiming that they verified the tokenizier was the same, which would imply it going at least beyond RL. Could still be a distill I think.
Looks like two Moonshot employees confirmed that it's not licensed before Moonshot made the decision to get out of the debate and delete their posts [0][1].
"Yes, Kimi K2.5 is an open source AI model. Developers and researchers can explore its architecture, build new solutions, and experiment openly. Model weights and code are publicly available on Hugging Face and the official GitHub repository."
Our only modification part is that, if the Software (or any derivative works
thereof) is used for any of your commercial products or services that have
more than 100 million monthly active users, or more than 20 million US dollars
(or equivalent in other currencies) in monthly revenue, you shall prominently
display "Kimi K2.5" on the user interface of such product or service.
Correct. (and I know you already know this but just for the record: (Nearly?) Everybody abuses the term "open source" when it comes to models. OSI have a post about it: https://opensource.org/ai/open-weights
Although it is not OSI approved, the license theoretically didn't add any more restrictions beyond attribution, which stays in line with The Open Source Definition.
Correct again -- CC- applies to data, not code -- weights are data, open weights suggests a creative commons approach …
“
CC-BY 4.0
Creative Commons Attribution 4.0 International
This license requires that reusers give credit to the creator. It allows reusers to distribute, remix, adapt, and build upon the material in any medium or format, even for commercial purposes.
BY
Credit must be given to you, the creator.
”
it's annoying the open source term is being cargo-culted around and I hate to say it but that ship looks like it has sailed.
funny that free software people were infuriated by the open source term and now the open source term is being completely misused in another context
Their definition matters more than most, I mean, anyone can define anything however they like. Hell, Windows is open-source, because I said so.
Also, even if it were not for the OSI, this still wouldn't be open source. Because there's no source code available. It's open-weight, which is a different thing. The models weights are, essentially, the "compiled" output. The input and algorithms, we don't know.
Cursor have said they are using Composer through their inference provider (Fireworks). Presumably the MIT is not viral like the GPL, so Cursor, and companies that use Cursor do not need to display Kimi attribution on their products.
It's definitely not what Kimi wanted, but it sounds like this is what is written.
"Just" Kimi K2.5 with RL—people really misunderstand how difficult it is to achieve these reults with RL. Cursor's research team is highly respected within the industry, and what they've done is quite impressive.
Before people go jumping to conclusions about model theft, it's worth considering the possibility that they did reach an agreement with Moonshot which their researchers were not aware of. That would certainly explain the deleted tweets. Until Moonshot makes an official statement, I'm not particularly concerned.
The amount of angst people feel the need to have against ai is incredible. We all seemed to want open weights, but it's time to take offense when open weights are used as intended?
Cursor is mostly an IDE / coding-agent harness company.
So it probably makes sense for them not to train their own base model, but instead license something like Kimi and fine-tune it for their own harness and workflows.
Their moat looks pretty thin. A VSCode fork with an open-source LLM fork on top. In the fast-moving coding-agent market, it’s not obvious they keep their massive valuation forever.
There is a plausible scenario in which software engineering requires a very finite amount of intelligence, in which sota models will be used mainly for other things and where for coding the harness will become increasingly more important than the model.
i've kinda had this thought before but never could express it ("you only need up to a certain level of smartness to express most coding concepts correctly")
but it never occurred to me that, if true, of course the harness becomes increasingly more important. which feels absolutely correct of course.
Their value is in the data they've collected and are collecting. Usage, acceptance rate, and all the connected signals. Plus having a large userbase where they can A / B test any finetune they create.
I hope for their sake they're using real metrics internally, and not whatever nonsense they're using to calculate stuff like "% written by LLM" in their dashboard, because that's... very wrong.
There are many reasons to make fun of Cursor. However, one of the things get right is their autocomplete model.
Are there any open models that come close? Why doesnt OAI or Anthropic dedicate some resources to blowing Cursor's model out of the water? Cursor's completion model is a sticking point for a lot of users.
I agree, their autocomplete (tab) model is the best, but recently I realised I am using it less and less - the new models are so good that I mostly just do agentic coding, and I do very little changes in the codebase by myself. This is probably a general trend and if the usage of autocomplete models is dying out, it's understandable the companies are not investing resources into it.
Don't get me started. For every half-decent choice, there's a multitude of insane choices. After all this time they still don't have side-by-side review.
Equally as annoying, the break from VSCode is horrible. Having to use a separate registry, not having basic settings sync, the delay behind mainline VSCode updates.
Then, it's just plain buggier than others. The agent terminal just doesn't work semi-regularly, it doesn't like listing directories in the @, the SSH plugin crashes every other time it tries to connect, undoing agent work undoes edits I made in unrelated files sometimes. Sometimes updates just regress performance hard for seemingly no reason.
I also noticed the token use is wildly less efficient than CC or Codex these days. After almost no time at all it's up to 100,000 tokens and they're charging $1 per request for Sonnet. Side-by-side, Cursor spent $17 in the same time CC spent $4. Which is bizarre to me, since they advertise how their indexing and semantic search is more token efficient?
The autocomplete model was the only reason I stayed as long as I did. I wish there was a VSCode equivalent.
Well, the UI as a whole is ok to me (except the parts which is way too volatile). I was talking about the UX of the autocomplete model. The model are very often spot on and fast, but it's impossible to properly configure it to be less in your face. Making it basically useless for day-to-day development.
Most companies don't do auto competition these days, including some that just recently stopped offering completion.
Which I find very unfortunate. There are so many cases, especially in proprietary codebases with non standard infrastructure, where good autocomplete is much better than "agentic" edits that produce nothing but slop which takes longer to clean up.
They’re pretty upfront in their release post that they took an open source model and improved it with their own coding data. They mention “continued pretraining” (on top of the base model) and RL. Cursor never claimed to have done a full pretraining run.
More to the point, beating Opus 4.6 at coding and coming within striking distance of gpt-5.4 is impressive! The benchmarks outperform raw Kimi K2.5.
It’s particularly impressive given larger labs like Meta are struggling to catch up to OpenAI/Anthropic.
This is exactly what Cursor should be doing, within the obvious bounds of the law and such. Not everyone needs a pristine foundation model. What a waste of compute. Anthropic & OpenAI need product-level competition to knock them off their $25/Mtok horse.
Indeed, this is quite obvious on Claude models vs Gemini. I fully believe Gemini is more powerful model, but the post training process is nowhere near what Anthropic does, which results in Gemini being horrible at coding sessions, while Claude is excellent.
At the same time, Moonshot violated Anthropic's ToS by training on their models' outputs :) [0]. And Anthropic violated copyright law by training on copyrighted material. It's violations all the way down.
There is no ToS at play here. There's only the license[1], which is MIT modified like so:
> Our only modification part is that, if the Software (or any derivative works
thereof) is used for any of your commercial products or services that have
more than 100 million monthly active users, or more than 20 million US dollars
(or equivalent in other currencies) in monthly revenue, you shall prominently
display "Kimi K2.5" on the user interface of such product or service.
Yes, this is pretty clear-cut. There's even a great alternative, namely GLM-5, that does not have such a clause (and other alternatives besides) so it feels a bit problematic that they would use Kimi 2.5 and then disregard that advertisement clause.
I've replied down the thread, but there are ways to go around that clause entirely, even if it would be enforceable. The obvious way is to have another company do the modification.
The worthwhile question AIUI is whether AI weights are even protected by human copyright. Note that firms whose "core" value is their proprietary AI weights don't even need this (at least AIUI) since they always can fall back on "they are clearly protected against misappropriation, like a trade secret". It becomes more interesting wrt. openly available AI models.
> The worthwhile question AIUI is whether AI weights are even protected by human copyright.
I'm also deeply curious about this legal question.
As I see it, model weights are the result of a mechanistic and lossy translation between training data and the final output weights. There is some human creativity involved, but that creativity is found exclusively in the model's code and training data, which are independently covered by copyright. Training is like a very expensive compilation process, and we have long-established that compiled artifacts are not distinct acts of creation.
In the case of a proprietary model like Kimi, copyright might survive based on 'special sauce' training like reinforcement learning – although that competes against the argument that pretraining on copyrighted data is 'fair use' transformation. However, I can't see a good argument that a model trained on a fully public domain dataset (with a genuinely open-source architecture) could support a copyright claim.
They probably licensed it. Still a bit deceptive not to mention it on the model card/blog post, but companies whitelabel all the time without mentioning.
It goes against the ML community ethos to obscure it, but is common branding practice.
Basically this is true for most startups in the world BUT Cursor, so here you are kinda inverting the logic of the matter. Cursor is at a size that, if they wanted to use K2.5, they could clearly state that it was K2.5 or get a license to avoid saying it.
IF we assume that the modified MIT clause is enforceable. And if we assume Cursor Inc. is running the modification. It could very well be the case that Cursor Research LTD is doing the modifications and re-licensing it to Cursor Inc. That would make any clause in the modified MIT moot.
The cursor investor pitch was we're training our own models to do coding. If your amazing model is just an RL repack, you need a new pitch to justify your 50bn valuation
Any investor who believed a team their size and with their capital was training a SOTA base model doesn't understand the space. I fully believe that was some of their investors, but people acting like RL + fine tuning based on their massive user base that's producing qualitatively better outputs than the base model is meaningless aren't understanding what the company is doing.
I think there's a reason why the people from Moonshot deleted their tweets; they're probably just researchers who got yelled at by the people who actually knew what was going on at Moonshot.
People need to seriously stop it with the whole reddit-esque Boston Marathon Bomber investigation-style low-info crusades. Its extremely unhealthy for both your own mental state and the state of discourse on the internet. Even if Cursor misbehaved (they did not): Your life is not materially changed whether they did or did not. Use it, or don't use it; these things are a matter that lies exclusively between Cursor and Moonshot.
From a users perspective, do we really care what model we're using under the hood? Or how well the software is solving our problems?
Seems like cursor is trying to build a "thicker wrapper" beyond the harness. Either to protect against Anthropic shutting them off or increase margins.
What do people like about cursor? I've been using it for the past couple days, and I just don't see many positive things about it. It seems people like the autocomplete so I'll have to give that a try.
There's just too many "features" the ux ends up being all over the place. I thought having the browser inside of the editor would be great for design, but it's not that much better than just having your browser open along with your editor.
I'm annoyed that we still don't know for certain which base model they used for Cursor 1.
This feels really rude to me. I have no problem with them fine-tuning open weight models to create their own - they are getting great results, and Cursor's research term should be respected for that. But deliberately hiding the base model they use is disrespectful of the researchers who created that model.
We used a Kimi base, with midtraining and RL on top. Going forward, we'll include the base used in our blog posts, that was a miss. Also, the license is through Fireworks: https://x.com/Kimi_Moonshot/status/2035074972943831491
As a paying customer, it just doesn't feel good that they are trying to pass off someone else's model as their own.
I mean I guess this is what businesses do all the time. There's a term for it even, it's called white-labeling.
But is this all that Cursor have? They pass of VS Code as their own, they pass off Kimi as their own... What do Cursor even do? What do I need them for?
Can you ensure that Notion is able to keep delivering given they don't develop their own models? Lovable? OpenCode? Should we be worried that Discord might disappear because they don't run their own data centers? Personally, I'm very concerned that one day Google might just have to close up shop, because while they do design their own chips, they don't fabricate them in-house; and don't get me started on TSMC and their critical dependency on ASML, they might as well just lock the doors.
White-labeling may be slightly dishonest to the consumer but the manufacturer and distributor are honest with each other. That doesn't appear to be the case here (Kimi's license requires publicly acknowledging Kimi is used for anyone operating at Cursor's scale).
cursor copy open source software repack it as closed source and made massive money. i don't want to hear anything on how Chinese is stealing and copying when the west is doing it themselves.
I noticed something strange with Cursor lately. When I am using Opus 4.6, sometimes it is giving ridiculously dumb answers as if they were actually using something like Qwen with a prompt to present itself as Opus. I have to close the session and start again hoping I'll get actual Opus.
There's no shot they're doing that. Would be suicide as soon as anyone notices, and by the looks of it, they didn't even clean up the URL here to "hide" the fact that this is Kimi K2.5 so i doubt there's any grand conspiracy here.
What's way more likely is that Opus has been quantized by anthropic or something similar. Or that Opus was updated and didn't work well with Cursor's harness after. Or a token caching issue. Etc.
I don't know - it works okay (yet to be tested whether it is actually smarter than Opus 4.6), but it is not bad at all. So far, it works quite fine (I'm not testing the "fast" version).
Honestly I don't think this leak is any good for Cursor. Not only this appears as a violation to Moonshot's ToS, this may also be in fact enough evidence for Anthropic to ban Cursor from using their models, just like they are doing to OpenCode.
Why? As I said before, Anthropic mentions Moonshot AI (Maker of the Kimi models) as one of the AI labs that were part of this alleged "distillation attack" [0] campaign and will use that reason to cut off Cursor, Just like they did to OpenAI, xAI and OpenCode.
Let's see if the market thinks Composor 2 is really that good without the Claude models helping Cursor. (If Anthropic cuts them off).
> this may also be in fact enough evidence for Anthropic to ban Cursor from using their models, just like they are doing to OpenCode.
The Anthropic ban on OpenCode isn't an Anthropic ban on OpenCode, it's a ban on using a Calude Code subscription with OpenCode. That's justified (or not) under various ToS arguments, but one can still use OpenCode with the more expensive API access.
Anthropic's complaint about distillation attacks is a distinct prong, one not levied against OpenCode. Additionally, the distillation activities described in your link don't describe Cursor's routine use of Anthropic's models. There, the model outputs are a primary product (e.g. the autocompleted code), and any learning signals provided are incidental.
Anthropic's complaint about "distillation" attacks (obligatory scare quotes because training on glorified chat logs is a far cry from actually distilling from model weights you have real access to) is also about ToS violations. Anthropic's ToS, like OpenAI's, forbids you from exploiting interactions with their model for the purpose of building a competitor, even though rumor has it that the AI industry has been doing exactly this for a long time anyway.
For all the muh productivity guys that like to claim they can turn invisible when no one is looking, an produce 600k lock over 6 weeks, well...cursor is useless now. We know kimi K2.5 won't make you 100 trillion times faster.
Incompetently repackaging. They started with VSCode so nearly all the work was already done, but still managed to make it leak memory like it's infinite. The power of AI slop! Their product is an anti-advertisement for the core concept of itself, which is kind of impressive.
I like Cursor's AI projects a lot. Cursor Tab is truly impressive. But you couldn't be more right.
I just downloaded VSCode again today after Cursor's latest update dropped my editor to 5 FPS or so (legitimately unusable. not hyperbole.) and holy shit it feels snappy. Completely forgot what it's like.
Ollama is also doing this.
There is so much money to be made repackaging open source these days.
So funny to see Twitter go wild saying "a 50 person team just beat Anthropic" blah blah.