> 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.
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.
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.
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).
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.
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.
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 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
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
"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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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).
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.
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.
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.
> 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.
It's all so predictable, even the comments here.
because its open source.
Like licenses are worth anything in the AI world…
A license doesn't matter if the perpetrator doesn't comply with it.
Open source licence requires attribution which obviously it is not done in this case.
No it doesn’t? Depends on the license
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.
> a 50 person team just beat Anthropic
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.
> packaging open source and reselling it.
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?
The meta data is useful.
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.
I guess they rely on many people not toggling privacy-mode on?
Does "code" include the prompt? Seems like the prompts would be the goldmines. Hook those up to rl an open weight model...
I doubt the majority does that. I bet the majority is using the defaults.
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).
As a command line junkie, what is the main thing Claude Code needs to catch up with cursor?
I haven't dove into using a LLM in my editor, so I am less familiar with workflows there.
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.
[1] https://marketplace.visualstudio.com/items?itemName=ezforo.c...
> There is so much money to be made repackaging open source these days
These days? Almost every tech offering in existence is 1000+ OSS dependencies gaffer taped together with a sprinkling of business logic.
Cursor isn't a shocking bit of software to pay for, its investment however...
The moat is the integration layer, not the model. I've seen this building MCP servers — structured data access matters more than which LLM you pick.
Do you know what Qwen model Composer 1.5 used?
> Cursor Composer 1 was Qwen...
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.
That's where it is going.
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].
[0] https://chainthink.cn/zh-CN/news/113784276696010804 - may have originally been https://x.com/apples_jimmy/status/2034920082602864990
[1] https://pbs.twimg.com/media/HD2Ky9jW4AAAe0Y?format=jpg&name=...
This is on their website...
"Is Kimi K2.5 open source?"
"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."
https://www.kimi.com/ai-models/kimi-k2-5
4th paragraph in license block
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.
My first reaction was "well, who knows how much revenue they're actually doing"
But at least the rumor mill has them significantly above that line:
> Revenue: As of March 2026, reports suggest Cursor has surpassed $2 billion in annualized revenue (ARR).
That's not an open source license, then.
Why not?
This 'Modified MIT' is not a license that has been through the OSI process: https://en.wikipedia.org/wiki/The_Open_Source_Definition#Com...
You can't just add random terms to an existing license and use its name. "Modified MIT: Like MIT but pay us 50 million dollars."
Perhaps CC-BY would've been more appropriate.
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
It wouldn't be regardless, because the model is open weights, not open source. It's just a license.
Which contradicts what they say on their website.
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
Another one, I believe this one was also deleted: https://x.com/HarveenChadha/status/2034933979720425611/photo...
TBH they really shouldn't have posted such a tweet in the first place, just sit back and watch their license enforced by the Internet.
I had the question "how do you even enforce this weird license term" back then, I guess I know the answer now.
"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.
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.
not sure if the hypothesis is even true though.
> Their moat looks pretty thin.
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.
That's every harness including VC Code Copilot.
People home about Teams sucking, but its market share is several times that of Slack because of distribution.
I guarantee that Microsoft has even more data.
There were conversations in the team yesterday about how Cursor's cloud agents are still ahead of Claude from a UX perspective.
Obviously we're running both, using the right tool for the job.
There is stickiness there from being early. That will be hard to replicate.
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.
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.
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.
Antigravity has an autocomplete model too. Based on Windsurf's, I guess.
Absolutely not. Windsurf also just stole an open source model, there’s almost zero chance Google is using that under the hood.
The model is great. The UX is ~~horrible~~ annoying...
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.
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.
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.
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.
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
https://www.bloomberg.com/news/articles/2026-03-12/ai-coding...
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.
What does this mean, that you can take Kimi and RL finetune it a little more and blow the big AI shops out of the water?
Would this have been extensively fine tuned, beyond what Anthropic/OAI would do themselves?
I guess this is Cursor's own benchmark, so you can finetune on your own dataset and get better results on your own specific tasks I guess.
Advertising your model with some obviously home grown benchmark is a bold play. It doesn't matter how good your model is, I immediately trust it less.
There's no "just" in RL. Fine tuning is very important and could make a lot of difference.
apparently GPT-5 uses the same pretrain as 4o did, hah
im pretty sure this is in violation of moonshot's ToS. this is going to be fun to watch unfold
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.
https://news.ycombinator.com/item?id=47126614
I thought Anthopoc’s training was deemed fair used. It was the downloading that was illegal
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.
[1] - https://huggingface.co/moonshotai/Kimi-K2.5/blob/main/LICENS...
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.
No they didn't [0][1]. With this leak they're probably negotiating as we speak, which could be why they've deleted the posts.
[0] https://chainthink.cn/zh-CN/news/113784276696010804
[1] https://pbs.twimg.com/media/HD2Ky9jW4AAAe0Y?format=jpg&name=...
I stand corrected, that is pretty scummy.
I bet Moonshot is going to make them open their wallets to avoid legal trouble.
Kimi K2.5 was released under a modified MIT license (100M+ MAU or $20M+ MRR has to prominently display Kimi K2.5). It will be fine.
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.
In practice nothing happens after violating an open source licenses, especially if you are willing to follow the terms after being notified.
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.
This is actually becoming a path dependency, a dependence on the supply chain.
Moonshot is raising money at a 10B usd valuation, cursor/anysphere is at a 30B usd valuation.
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.
Simon, sorry to hijack the thread, but what is a good way of contacting you? I'd love to pick your brain on an AI talk I'm supposed to be giving soon.
To be fair, is "with RL", "just"?
They should have disclosed it though. If they didn't it's a bad look for sure.
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.
Scores higher than Opus 4.6 on their in-house benchmark? Sounds legit.
A question. I’m due for a yearly Cursor subscription renewal, how does the credit limit look like?
Currently I’ve not hit any of the limits despite using it quite rigorously, I wonder if this will change with a renewal?
But you have to buy into it right? If you don't have a limit then what did your contract look like?
I signed up last year when the limits were a lot more generous.
Honestly, this is pretty much how most of the new models operate nowadays: a base model combined with RL and some product-layer magic.
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?
As a paying customer, I don't care where the model comes from, I only care how good it is.
Sure, and also at what price point.
But can I rely on Cursor to be able to keep delivering, when they aren't the one's doing the work themselves?
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.
Well, they can keep stealing as long as someone open weight their models.
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).
YC is back at it again.
This whole ai stuff feels like a big bubble especially with the oil price soon at $200 and guaranteed recession.
is this the model used on free mode?
Their first model was also based on an open source Chinese base model. They never fully trained their own model.
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).
I whish it was GLM 5.0.
Cursor can't compete with Claude Code's subsidized pricing, so they are trying to gaslight people that their cheap model is good enough.
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).
[0] https://www.anthropic.com/news/detecting-and-preventing-dist...
> 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.
Kimi K2.5 is an open source model. It is intended for people to make derivative models.
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.
Cursor is killed for this market.
A hyped startup providing zero added value, burning investor money only to repackage somebody else's work? That's new... /s
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.
It depends on what you consider value. People are using it so they find some value.