Claude Distillation Claims Return After Researchers Probe Kimi K3

Kimi K3 and Claude distillation are back under scrutiny after new research found unusual similarities between Moonshot AI’s model and Anthropic’s Claude reasoning patterns.

Back in February, Anthropic accused Chinese AI companies of trying to copy its AI models. A few months earlier, Anthropic accused Alibaba directly of attempting to copy Claude AI.

Alibaba later banned Claude Code from its workplace.

Researchers find a way to expose hidden AI reasoning

A new research paper has reignited debate over whether Moonshot AI’s Kimi K3 could have been distilled from Anthropic’s Claude models.

Critics previously questioned that possibility because only about two weeks separated the release of Anthropic’s Fable and Kimi K3. That appeared to leave little time for large-scale distillation.

Researchers now say an API-related weakness could have made reasoning-trace extraction possible without breaking the underlying encryption.

AI companies generally hide internal reasoning while transferring encrypted reasoning blocks between models. Researchers found that systems from OpenAI, Anthropic, and Google could allow those blocks to work across different models within the same ecosystem.

That means reasoning generated by a stronger model could potentially reach a weaker sibling model capable of revealing information from the encrypted block.

Kimi K3 shows unusual compatibility with Claude reasoning

Researchers tested Kimi K3 using a small section of decoded Claude Opus 4.8 reasoning.

According to the paper, supplying only 1% of the Claude reasoning fragment significantly changed Kimi K3’s later reasoning and pushed its wording and behavior closer to Claude.

Researchers also found that some Claude and GPT reasoning sequences were considerably easier to extract from Kimi K3 than from competing open models.

For certain reasoning spans, the paper reports that extraction was roughly six orders of magnitude easier from Kimi K3 than from the next-closest model.

Kimi K3 also showed an unusually high probability of predicting and continuing Claude-generated text compared with models including DeepSeek-V4-Flash.

The findings do not prove Kimi K3 was distilled from Claude

The researchers argue that Kimi K3 demonstrates unusually strong compatibility with Claude’s reasoning patterns.

Those results could support the theory that some form of distillation occurred, but they do not establish how Moonshot AI trained Kimi K3.

Without access to Moonshot’s training data and methodology, researchers cannot conclusively verify the allegations.

The findings therefore add circumstantial evidence to the broader Claude distillation debate rather than proving that Moonshot copied Anthropic’s models.

In other AI news, Claude AI text will embed hidden watermarks and metadata in the EU.

Via Wccftech