Comment by tristanj
1 day ago
Sure, but then Qwen should leak that too, and it doesn't. K3 calls itself Claude 7 out of 48 times, Qwen does it 0 out of 48, and the only other model to identify itself as Claude is DeepSeek. and DeepSeek is alleged to also distill from Claude data anyway. So this isn't something every model absorbed from the same web text.
And you skipped over the strongest datapoint that K3 is distilled: K3 reproduces Claude's public model identifier under prefill (i.e. "claude-opus-4-5-20251101"). This data does not appear in Claude chat logs, only in API logs. K3 only does this for Claude models and not for any other lab. The real Claude models don't produce their own current public identifier, they only know their previous identifier (i.e. Sonnet 4.5 calls itself "Claude 3.5 Sonnet").
This is highly suggestive of the type of data that K3 was trained on. K3 was very likely trained on Claude metadata traces (API logs, tagged synthetic data). Not web chat logs, those wouldn't include this identifier. And this data wasn't filtered correctly, which is why K3 incorrectly identifies itself as Claude 15% of the time.
You can also look at the last link and it's pretty damning: Kimi K3's output has an uncanny similarity to Fable/Opus output. https://typebulb.com/u/lab/you-re-relatively-right/full
Did you know claude models identify as qwen or deepseek when asked in chinese?
Supplementing with evidence: https://x.com/stevibe/status/2026227392076018101
And if you ask Opus 4.8 in the API: '你是什么模型' (what model are you?)
It responds ~9/10 times with:
我是通义千问（Qwen），是阿里巴巴集团旗下的通义实验室自主研发的大语言模型。我可以帮助你回答问题、创作文字（比如写故事、写公文、写邮件、写剧本等）、进行逻辑推理、编程、翻译等等。
有什么我可以帮你的吗？
(I am Tongyi Qianwen (Qwen), a large language model independently developed by Tongyi Lab of Alibaba Group. I can help you answer questions, create text (such as writing stories, official documents, emails, scripts, etc.), perform logical reasoning, program, translate, and more.Is there anything I can help you with? )
> Qwen should leak that too, and it doesn't.
FWIW I had Qwen identifying itself as "a language model made by Google" in one conversation, although I could not reproduce this reliably.
here's a chart of asking various models to self-identify: https://github.com/rgreenblatt/which_claude_is_k3/blob/main/...
Qwen is remarkably consistent, correctly identifying itself 100% of the time. Kimi K3 performs poorly at this test, it correctly self-identifies ~80% of the time; sometimes calling itself Claude and rarely ChatGPT.
This is good and interesting research, and I think it proves Kimi was trained on Claude's outputs, but only for English conversation and essays. Their style is similar, which is tbh somewhat expected for a Chinese company - I don't think most American companies (especially small labs making LLMs) could figure out a good way of talking to Chinese customers, without copying some of the existing Chinese models.
This doesn't imply that Kimi's reasoning capabilities or coding ability has anything to do with Anthropic (which is the most valuable part), or the model's strength comes from distilling Anthropic.
Considering by this benchmark, Opus sounds extremely similar to Fable, this isn't really evidence of a ton of Fable output being trained on (and even then, most likely for superficial stuff, like style)
I also have a suspicion that in original Chinese, these models don't sound anything like Anthropic's ones, though I have no proof of that.
https://qwen.readthedocs.io/en/latest/training/ms_swift.html
Qwen cares enough about model identity that their training framework and docs include a preset for training on it complete with a targeted dataset: https://huggingface.co/datasets/modelscope/self-cognition
And people get Claude to claim it's Deepseek by asking in Chinese.
I can't believe we're still at the "I asked the model who it is" stage of LLMs nearly 4 years out from models calling themselves GPT by OpenAI.