The accusation that Moonshot AI distilled Anthropic's Fable model to create Kimi K3 has circulated widely, but a close examination of the technical requirements and timeline suggests the claim rests on shaky ground. When you understand what distillation actually requires, the numbers Anthropic published don't support the story being told.

What Exactly Is Model Distillation, and Why Does It Matter?

Model distillation is a technique where a smaller "student" model learns to replicate the behavior of a larger "teacher" model. The concept dates back to 2015 research by Geoffrey Hinton and colleagues. But distillation comes in several flavors, each requiring different levels of access to the teacher model.

The most powerful form of distillation requires access to the teacher model's internal probability scores, known as logits. These are the raw numerical scores the model assigns to every possible next word it could generate. A student model trained on these logits can match the teacher's decision-making process with remarkable efficiency. For example, Qwen used logit-level distillation to produce six of its eight Qwen3 models from just two flagship models, using roughly one-tenth the computing power that training from scratch would require.

The next tier down is on-policy distillation, where the student generates text and the teacher scores each token the student produces. This is now standard practice at the frontier of AI development. DeepSeek-V4, for instance, trains separate reinforcement learning experts for different domains and then distills them into a unified student model.

But here's the critical issue: proprietary APIs like Anthropic's hide these internal scores entirely. Anthropic does not provide logits through its API, and it returns only a summary of the model's reasoning rather than the raw chain of thought. This means the most effective forms of distillation are simply not possible through normal API access.

Can You Distill a Model Using Only What You Can See?

The lowest tier of distillation uses only visible text. This is what DeepSeek did when it created its distilled reasoning models. DeepSeek trained six open-weight models on roughly 800,000 sampled reasoning traces from its R1 model using straightforward supervised fine-tuning. The result was striking: DeepSeek-R1-Distill-Qwen-32B scored 72.6 on the AIME 2024 mathematics benchmark, compared to just 9.3 for GPT-4o. While the distilled student still fell short of R1 itself at 79.8, it recovered about 91 percent of the teacher's performance on that benchmark.

This approach requires access to the teacher model's actual reasoning traces. But Anthropic built a specific safety classifier to prevent exactly this kind of extraction. The classifier, documented in Anthropic's cookbook for Fable 5, includes a category called "reasoning_extraction" designed to catch requests attempting to extract the model's summarized thinking. When triggered, the request reroutes to a different model, Opus 4.8, meaning some harvested data would not actually be from Fable.

How to Evaluate Claims About AI Model Theft

• Check the Timeline: Anthropic published its evidence of distillation attempts on February 23, but Fable 5 did not launch until June 9. Kimi K3 shipped on July 16, giving only 18 days of potential data collection after Fable became available. Three weeks is not impossible, but it is tight for the scale of training described.
• Verify Technical Feasibility: Determine what level of access the accused party would need. If the claim requires logits or internal probability scores, and the API does not provide them, the accusation becomes much less plausible. Anthropic's API provides neither logits nor raw reasoning traces.
• Examine the Evidence: Look for concrete proof like access logs, training data indicators, or procurement documents. The White House claim disclosed none of these. The only evidence cited was Anthropic's report documenting 3.4 million Claude exchanges attributed to Moonshot, but this does not prove those exchanges were used for distillation.

The pattern of accusation is worth noting. Similar claims emerged in January 2025 when DeepSeek released a competitive model, and they arrive reliably within days of a Chinese AI lab releasing a product that competes with US models. The accusation is always framed as theft.

Why the Fable Numbers Don't Support the Distillation Story

Anthropic's February 23 report documented over 16 million Claude exchanges it attributed to distillation campaigns, with 3.4 million of them traced to Moonshot. But the report predates Fable by more than three months. When Kimi K3 was published on July 16, Fable had been publicly available for only 18 days. Even accounting for the speed of modern data collection and training, the timeline creates a credibility gap.

Moreover, the technical requirements for distillation at scale are substantial. To recover 91 percent of a teacher model's performance through reasoning trace distillation, as DeepSeek did, requires not just access to the model but the ability to extract and process hundreds of thousands of reasoning traces. Anthropic's anti-distillation classifier was specifically designed to prevent this extraction. A June 2026 study found that elicited reasoning traces delivered 96.7 percent of the value of true internal access, but hiding the trace still costs roughly 3 percent in performance. This is a meaningful penalty that would show up in benchmarks.

The broader context matters too. The appearance of open-weight frontier models means there is no longer a need to pay premium prices for proprietary APIs to build competitive systems. AI costs can be significantly reduced through compliant and data-safe methods. This shift is reshaping the economics of AI development, and the panic around distillation accusations may reflect anxiety about this transition rather than evidence of actual theft.

The story of Kimi K3 is not primarily about distillation. It is about who captures the margin on inference, the cost of running AI models at scale. The technical claims are being selected for effect rather than derived from evidence. Once you examine what distillation actually requires and what Anthropic's own published numbers show, the narrative becomes much harder to sustain.