China’s Kimi K3 Calls Itself Claude, Exposing Illegal Distillation

Moonshot AI’s new Kimi K3 model has quickly become one of the most talked-about AI releases of the week after topping a major front-end coding benchmark.

The Chinese AI company says Kimi K3 is a 2.8 trillion-parameter open model with native multimodal abilities and a 1 million-token context window. Moonshot describes it as its most capable model yet, built for long coding tasks, knowledge work, and reasoning.

A Major Jump in Coding Rankings

Kimi K3 drew attention after reaching the top spot on Arena’s Frontend Code ranking.

According to Tom’s Hardware, Arena ranked Kimi K3 first with 1,679 points, ahead of Anthropic’s Claude Fable 5 in blind developer testing. The result marked a sharp jump from Kimi K2.6, which had reportedly been ranked 18th before Kimi K3 moved to first place.

Moonshot says Kimi K3 uses Kimi Delta Attention and Attention Residuals, two architecture changes designed to improve performance at long context lengths and large model scale. The company also claims these changes deliver around a 2.5x improvement in scaling efficiency compared with Kimi K2.

Distillation Claims Surface Online

The benchmark win has also triggered questions about how Moonshot trained the model.

Wccftech reported that some online users and analysts believe Kimi K3 may have been distilled from Anthropic’s Claude models. The report pointed to anecdotal evidence, including at least one shared conversation in which Kimi K3 allegedly identified itself as “Claude, an AI assistant made by Anthropic.”

The report also argued that Kimi K3’s training and inference efficiency appear unusual. It claimed that if the model was trained with significantly less compute than frontier U.S. models, that efficiency should also be more visible in inference costs. Wccftech said the mismatch has added to speculation that distillation may have played a role.

No Public Proof Yet

However, the claims remain unproven.

A model identifying itself as another AI system can raise questions, but it is not definitive proof of distillation. Such behavior can also come from training data contamination, system prompt issues, roleplay leakage, or copied examples from public datasets.

Wccftech also described the matter as “educated conjecture,” not a confirmed finding. Moonshot has not publicly confirmed that Kimi K3 was trained on outputs from Anthropic models.

Moonshot Points to Its Own Architecture

Moonshot’s official explanation focuses on model architecture and training improvements.

The company says Kimi K3 activates 16 of 896 experts through a sparse mixture-of-experts design and uses Stable LatentMoE, Kimi Delta Attention, and Attention Residuals to improve scaling. It also says Kimi K3 applies quantization-aware training and is designed for deployment on large accelerator clusters.

Moonshot says the model is available through Kimi.com, Kimi Work, Kimi Code, and the Kimi API. It also plans to release the full model weights by July 27, 2026.

Debate Over AI Distillation Continues

The Kimi K3 controversy comes as the AI industry continues to debate model distillation.

Distillation generally involves using outputs from a stronger model to train or improve another model. It is common in some forms of AI research, but it becomes controversial when companies are accused of using closed commercial models to train competing systems without permission.

For now, Kimi K3’s benchmark performance is real, but the claim that it was distilled from Claude remains an allegation. Until more technical evidence or the full model report becomes available, the safest conclusion is that Moonshot has delivered a major AI benchmark result while also attracting serious questions about how that result was achieved.