This data sort of disqualifies itself: unless Moonshot has a time machine, K3 should be more similar to Opus 4.5-4.8 than Fable 5.
Keep in mind, Anthropic started limiting access and introduced anti-distillation measures around 4.5-4.6 (?). So the majority of distillation should have happened on earlier models.
Maybe a better explanation is that they have access to the same training datasets? Which if private can again raise questions about theft, but on a very different level.
I also suspect the questions asked matter a lot, and the system prompts matter a lot, because "the map is built from nothing but the words they choose" - so this is more a measure of linguistic style than anything.
If you use the Claude Code harness on two models you will probably get very similarly styled output. I would not be surprised if K3 "stole" a lot of the harness that was leaked.
Notice how the only rows that even come close to being gray is GPT-5.4+Mini, diverging even from other GPT models. Is this because it has a wholly different training set, or (more likely IMO), did it just have a system prompt that leads to a different style?
Worthless. Any LLM output similarity metric that uses n-grams as a source might as well measure the average temperature on Mars surface, no matter how much lipstick you put on it. It just can't have enough certainty. There used to be an n-gram benchmark popular on Twitter that showed extreme similarity of grok-3-beta to gpt-4.5-preview, while these models were trained on new base ones, came out 2 weeks apart, and were unmistakably different. Results were wildly inconsistent run to run. It didn't stop the crowd believing its creator in that DeepSeek R1 was trained on o1-preview (which was obvious bullshit as well, they were as different as two models can be). It's amazing how you can put anything on the web and everybody will believe you without checking or even understanding of what they're looking at.
K3 was trained on Claude's outputs, though - it repeats Anthropic's prompt injections 1:1 in its reasoning, which you should know if you ever tinkered with both models long enough. Good for them.
All frontier models have been trained without any regards for IP protection laws. I don't see how anyone can argue in good faith that distillation is not fair game and does not ultimately "benefit humanity™"
I get what you're saying and two wrongs do not make a right but the irony, and why people are even talking about this, is that the thing being distilled clearly, and knowingly, violated copyright & terms across the entire internet.
For me, I don't really care about the theft aspect but when people are claiming that these open models are better value or going to overtake anthropic/openai models, the implication that the open models are training of distilled data means all the "progress" they are making is just mimiced from the closed models.
It's a bit interesting how the open models are able to keep pace with the closed models except whole maintaining a steady following time.
It's important to note that even if these open models are distilled, they are showing genuine improvements in their architecture, which enables inference costs to be several factors below what equivalent closed models have.
The interesting question is: will Anthropic release a Fable like model with an architecture similar to Kimi, and get the inference cost gains? They should surely beat Kimi because they can internally distill as much as they want.
Distillation is absolutely not the reason they're good. It's not necessarily even done on a more capable model. It can even be done on itself and still bring improvement, or on a weaker model as well (see GLM and Gemini, which is definitely true because it repeats Deepmind's injections).
Training a model is very expensive and creates something no individual rights-holder could. Distilling a model copies this value add and captures it without bearing the cost that created it.
I understand you think Anthropic should have paid for the information it trained the models on. But im talking about all the costs to build a model. Do you think Anthropic didnt spend money to build these models? Did you not know that its actually very expensive?
This actually undermines the argument that distilling is harmless because its founded on the idea that Anthropic did the same thing and didn’t have any repercussions.
They only had to pay for storing the books on a server for later possible use. They did not have to pay anything for the training which was declared fair use.
Could you give an example of the value that only training a model can create but none of the rights-holder could? I feel like if you got a direct, instant communication channel to any of the rights-holder that created the content in the training set of those models, you'd get more value than what the LLM could ever give you on any specific subject.
the outputs of the model have no property protections, and training a model on the outputs of another model does the exact same thing - its expensive and creates new value over what was in the input - a set of documents.
given that github is full of vibe coded repos, reddit is full of bot comments, and most blog posts are ai slop, isnt it guaranteed that anyone training on big public data sets will be closely tracking each other in text style?
If I am working on something not sensitive and afterwards upload my chat log to a public dataset, can open models use that data for training?
Keep in mind, I paid for access to model and half of the work is mine (the prompts). But can the model provider forbid me from publishing my result as open data?
I didn’t participate in the discussion yesterday because I find it implausible Fable was available long enough (3-4 weeks cumulative?) to get data and train and have it fundamentally affect.
But I don’t grok training enough to know that’s silly.
If my new prior is you can…that’s a pretty thin moat that’s essentially indefensible.
My guess is they used the other Anthropic models extensively for synthetic data generation. The top most similar models for K3 are (lower means more similar)
Fable 5 -> 0.42
Opus 4.8 -> 0.45
Sonnet 5 -> 0.45
Opus 4.7 -> 0.46
Grok 4.3 -> 0.52
There's an obvious jump at Grok 4.3, and it would not surprise me that the similarity there is because Grok used Anthropic models for training too (you can get the similarity list for Grok and it does look like the top most similar models for Grok are either Anthropic models or some Chinese models).
The damning evidence that K3 used Anthropic models for training is that K3 is more similar to those models than it is to K2.6. If you look at the Anthropic, OpenAI or Google models, they are most similar with their own other models. Not so with K3, where K2.6 is less similar than 15 other models.
Now, why is Fable 5 the most similar to K3 and not Opus 4.8. I think it's quite likely that K3 did some fine tuning at the end, when Fable 5 became available. They probably had all the infrastructure in place, and Mythos had been announced for months, so they were probably waiting for the second the newest Anthropic model was released to start using it for synthetic data generation.
Note that until last generation all Chinese models were relatively small. This was mainly due to lack of training hardware, but as soon as the new Huawei NPUs started shipping some Chinese labs switched to larger models:
Deepseek: 670B to 1.6T
Moonshot: 1.1T to 2.8T
Xiaomi: 310B to 1T
The new models also use a different architecture so I would assume in tests they will look different from previous generations.
Z.ai (GLM) and MiniMax on the other hand have continued training existing models (with some surgical changes to improve long context memory) so they should score similarly in these tests.
This looks very similar to the claim that distilling a model from Anthropic is the same thing as Anthropic distilling the model from information on the internet.
Which is very flawed, since distillation requires the thing to exist in the thing it’s distilled from. And no LLM model existed in the information Anthropic used to train the model. Instead the model was built using information and utilizing new technology including hardware, software, transformer architecture, etc.
I don't believe I understand your argument? Are you claiming a moral, legal or practical difference? Or are you saying that Anthropic spending resources on training an LLM is somehow different from an author spending resources on writing a book?
In any case, I doubt Kimi was trained without "stealing" the same data. Assembling all of your training data from Claude responses seems infeasible. It's much more likely that Kimi's base model was trained similarly to any other base model, with terabytes of data from all imaginable sources. Then the model was fine-tuned with "high-quality" data, followed by reinforcement learning. Throwing in lots of chat transcripts from other chatbots into the "high-quality" dataset would be expected, and is done to some degree by everyone, but maybe a lot more for Kimi. And likely they did a lot of reinforcement learning against the Claude API
The model would exist without Claude, it just wouldn't be nearly as coherent or smart
Think of it in terms of distilled knowledge, not distilled LLMs.
I find both claims unsound, though. Knowledge or model behavior itself is not copyrightable, so all these claims just boil down to the "I am not happy with that" argument. You cannot claim someone is stealing something you don't own in the first place.
You misunderstand the point. You can think both or either are morally right/wrong or good/bad for society. But distilling a model and building one using information online are fundamentally different. Even of you think the information the frontier models used was not fairly accessed.
You might as well say a photocopy of a book was "built using information and utilizing new technology including hardware, software, transformer architecture, etc."
All ML/AI models are comparable to some form of compression(al beit lossy) of information and in this case copyrighted information. The OP is pointing to this as stolen data(by all the companies that started with pre trained models)
Maybe both are ok. Maybe neither. Maybe one. The point is this does not follow
>If training on copyrighted data without authors consent is ok then distilling is ok as well.
Because building a model from distillation and building a model from raw data are not the same. You have to evaluate them independently. And legally its different as well. IP vs ToS (civil).
This data sort of disqualifies itself: unless Moonshot has a time machine, K3 should be more similar to Opus 4.5-4.8 than Fable 5.
Keep in mind, Anthropic started limiting access and introduced anti-distillation measures around 4.5-4.6 (?). So the majority of distillation should have happened on earlier models.
Maybe a better explanation is that they have access to the same training datasets? Which if private can again raise questions about theft, but on a very different level.
So this shows distance relative to other models, but I don't have a good sense for what these numbers say in absolute terms.
K3-to-Fable is blue at 0.42. Is 0.42 meaningful, or did we set 0.4 as the lower bound because it makes 0.42 look significant?
Sol-to-Fable is 0.69. It's dark yellow, making this look VERY different from 0.42. But is it? What do these numbers mean in absolute terms?
I also suspect the questions asked matter a lot, and the system prompts matter a lot, because "the map is built from nothing but the words they choose" - so this is more a measure of linguistic style than anything.
If you use the Claude Code harness on two models you will probably get very similarly styled output. I would not be surprised if K3 "stole" a lot of the harness that was leaked.
Notice how the only rows that even come close to being gray is GPT-5.4+Mini, diverging even from other GPT models. Is this because it has a wholly different training set, or (more likely IMO), did it just have a system prompt that leads to a different style?
I think it needs some more explanation - Fable is 0.42 from itself apparently so Kimi K3 is basically indistinguishable?
>character trigrams
Worthless. Any LLM output similarity metric that uses n-grams as a source might as well measure the average temperature on Mars surface, no matter how much lipstick you put on it. It just can't have enough certainty. There used to be an n-gram benchmark popular on Twitter that showed extreme similarity of grok-3-beta to gpt-4.5-preview, while these models were trained on new base ones, came out 2 weeks apart, and were unmistakably different. Results were wildly inconsistent run to run. It didn't stop the crowd believing its creator in that DeepSeek R1 was trained on o1-preview (which was obvious bullshit as well, they were as different as two models can be). It's amazing how you can put anything on the web and everybody will believe you without checking or even understanding of what they're looking at.
K3 was trained on Claude's outputs, though - it repeats Anthropic's prompt injections 1:1 in its reasoning, which you should know if you ever tinkered with both models long enough. Good for them.
This looks to be more behavior based, not logit based? What is the actual claim?
It seems to boil down to China bad, US good, but with tech people speculating so it's definitely valid
honestly 99% of hn, reddit, etc. is the exact opposite of that statement.
Keep in mind if you ask Claude what model it is in Chinese it says its Deepseek or Qwen or Kimi.
So who's training on who's outputs?
https://news.ycombinator.com/item?id=48990086
https://x.com/stevibe/status/2026227392076018101
How do you do cross entropy analysis for Claude without logits?
I assume Kimi does provide logits. You can do cross entropy both ways. With slightly different results.
Edit: No that doesn't seem to be what's happening here. I think it's some kind of word frequency analysis.
Shouldn’t we expect a convergence as models become more powerful ?
There is an optimal answer to any question. Something that maximizes utility and minimizes tokens.
Honest take: different surface, same category.
All frontier models have been trained without any regards for IP protection laws. I don't see how anyone can argue in good faith that distillation is not fair game and does not ultimately "benefit humanity™"
Does distilling actually violate any IP protection laws? Sure, it’s against their terms of service.
I get what you're saying and two wrongs do not make a right but the irony, and why people are even talking about this, is that the thing being distilled clearly, and knowingly, violated copyright & terms across the entire internet.
Yeah, everyone is saying that but I don't find the irony very interesting.
If it does then their original training did as well.
One mentions distillation to intimate that one is in fact still the leader. It has zero to do with fairness.
For me, I don't really care about the theft aspect but when people are claiming that these open models are better value or going to overtake anthropic/openai models, the implication that the open models are training of distilled data means all the "progress" they are making is just mimiced from the closed models.
It's a bit interesting how the open models are able to keep pace with the closed models except whole maintaining a steady following time.
It's important to note that even if these open models are distilled, they are showing genuine improvements in their architecture, which enables inference costs to be several factors below what equivalent closed models have.
The interesting question is: will Anthropic release a Fable like model with an architecture similar to Kimi, and get the inference cost gains? They should surely beat Kimi because they can internally distill as much as they want.
Distillation is absolutely not the reason they're good. It's not necessarily even done on a more capable model. It can even be done on itself and still bring improvement, or on a weaker model as well (see GLM and Gemini, which is definitely true because it repeats Deepmind's injections).
Training a model is very expensive and creates something no individual rights-holder could. Distilling a model copies this value add and captures it without bearing the cost that created it.
Yes. I bet Moonshot paid for API access as opposed to pirating like Anthropic did.
I understand you think Anthropic should have paid for the information it trained the models on. But im talking about all the costs to build a model. Do you think Anthropic didnt spend money to build these models? Did you not know that its actually very expensive?
What of the costs for creating the data that was used to train the model being distilled?
At least 1.5B by stealing books, per a recent ruling.
I don’t feel sorry for the model companies
This actually undermines the argument that distilling is harmless because its founded on the idea that Anthropic did the same thing and didn’t have any repercussions.
They only had to pay for storing the books on a server for later possible use. They did not have to pay anything for the training which was declared fair use.
Do you people seriously believe that Moonshot and the Chinese personal-cult-state dont possess and train on the same torrents?
Well of course. But they dont hate Chinese model companies.
They aren’t hypocritically crying foul about it, so no we don’t care if they do.
Could you give an example of the value that only training a model can create but none of the rights-holder could? I feel like if you got a direct, instant communication channel to any of the rights-holder that created the content in the training set of those models, you'd get more value than what the LLM could ever give you on any specific subject.
>Could you give an example of the value that only training a model can create but none of the rights-holder could?
Yeah.
LLMs
the outputs of the model have no property protections, and training a model on the outputs of another model does the exact same thing - its expensive and creates new value over what was in the input - a set of documents.
I'm going to hope this was sarcasm and if it is, it's great.
An incredibly ironic comment.
There is a lot more one has to be good at to make a fable level model. Distillation won't get you there, it will help refine some at the end
A lot of those arguments could be said about writing a book, or a decent forum guide.
given that github is full of vibe coded repos, reddit is full of bot comments, and most blog posts are ai slop, isnt it guaranteed that anyone training on big public data sets will be closely tracking each other in text style?
I would gladly make a donation of Claude accounts to Kimi, to support more distillation.
On a more serious note:
If I am working on something not sensitive and afterwards upload my chat log to a public dataset, can open models use that data for training?
Keep in mind, I paid for access to model and half of the work is mine (the prompts). But can the model provider forbid me from publishing my result as open data?
AI stealing from AI? That's shocking that an industry built on stealing would engage in stealing.
I didn’t participate in the discussion yesterday because I find it implausible Fable was available long enough (3-4 weeks cumulative?) to get data and train and have it fundamentally affect.
But I don’t grok training enough to know that’s silly.
If my new prior is you can…that’s a pretty thin moat that’s essentially indefensible.
My guess is they used the other Anthropic models extensively for synthetic data generation. The top most similar models for K3 are (lower means more similar)
Fable 5 -> 0.42
Opus 4.8 -> 0.45
Sonnet 5 -> 0.45
Opus 4.7 -> 0.46
Grok 4.3 -> 0.52
There's an obvious jump at Grok 4.3, and it would not surprise me that the similarity there is because Grok used Anthropic models for training too (you can get the similarity list for Grok and it does look like the top most similar models for Grok are either Anthropic models or some Chinese models).
The damning evidence that K3 used Anthropic models for training is that K3 is more similar to those models than it is to K2.6. If you look at the Anthropic, OpenAI or Google models, they are most similar with their own other models. Not so with K3, where K2.6 is less similar than 15 other models.
Now, why is Fable 5 the most similar to K3 and not Opus 4.8. I think it's quite likely that K3 did some fine tuning at the end, when Fable 5 became available. They probably had all the infrastructure in place, and Mythos had been announced for months, so they were probably waiting for the second the newest Anthropic model was released to start using it for synthetic data generation.
Note that until last generation all Chinese models were relatively small. This was mainly due to lack of training hardware, but as soon as the new Huawei NPUs started shipping some Chinese labs switched to larger models:
Deepseek: 670B to 1.6T
Moonshot: 1.1T to 2.8T
Xiaomi: 310B to 1T
The new models also use a different architecture so I would assume in tests they will look different from previous generations.
Z.ai (GLM) and MiniMax on the other hand have continued training existing models (with some surgical changes to improve long context memory) so they should score similarly in these tests.
Stolen data was stolen. Oh no! Anyway.
Can you elaborate?
This looks very similar to the claim that distilling a model from Anthropic is the same thing as Anthropic distilling the model from information on the internet.
Which is very flawed, since distillation requires the thing to exist in the thing it’s distilled from. And no LLM model existed in the information Anthropic used to train the model. Instead the model was built using information and utilizing new technology including hardware, software, transformer architecture, etc.
I don't believe I understand your argument? Are you claiming a moral, legal or practical difference? Or are you saying that Anthropic spending resources on training an LLM is somehow different from an author spending resources on writing a book?
In any case, I doubt Kimi was trained without "stealing" the same data. Assembling all of your training data from Claude responses seems infeasible. It's much more likely that Kimi's base model was trained similarly to any other base model, with terabytes of data from all imaginable sources. Then the model was fine-tuned with "high-quality" data, followed by reinforcement learning. Throwing in lots of chat transcripts from other chatbots into the "high-quality" dataset would be expected, and is done to some degree by everyone, but maybe a lot more for Kimi. And likely they did a lot of reinforcement learning against the Claude API
The model would exist without Claude, it just wouldn't be nearly as coherent or smart
Think of it in terms of distilled knowledge, not distilled LLMs.
I find both claims unsound, though. Knowledge or model behavior itself is not copyrightable, so all these claims just boil down to the "I am not happy with that" argument. You cannot claim someone is stealing something you don't own in the first place.
You misunderstand the point. You can think both or either are morally right/wrong or good/bad for society. But distilling a model and building one using information online are fundamentally different. Even of you think the information the frontier models used was not fairly accessed.
You might as well say a photocopy of a book was "built using information and utilizing new technology including hardware, software, transformer architecture, etc."
Yeah but thats to my point. You can’t legally resell a photocopy of a book.
All ML/AI models are comparable to some form of compression(al beit lossy) of information and in this case copyrighted information. The OP is pointing to this as stolen data(by all the companies that started with pre trained models)
Models being compression algorithms doesn’t really make your point. Compression algorithms are copyrightable.
So what?
If training on copyrighted data without authors consent is ok then distilling is ok as well.
Maybe both are ok. Maybe neither. Maybe one. The point is this does not follow
>If training on copyrighted data without authors consent is ok then distilling is ok as well.
Because building a model from distillation and building a model from raw data are not the same. You have to evaluate them independently. And legally its different as well. IP vs ToS (civil).