Kimi K3 и Fable от Anthropic: эксперты сомневаются в версии о дистилляции

TechCrunch reported on July 23, 2026, citing experts who believe the strength of the new Kimi K3 model is not explained by "distillation" of Anthropic's Fable model, as some observers had suggested, but by more serious engineering work on the part of its creators.

What the experts said

Experts polled by TechCrunch on July 23, 2026, are questioning the theory that Kimi K3 became powerful by copying answers from Fable — Anthropic's flagship model. In their assessment, distillation alone is not enough to produce a model of this class, especially in such a short timeframe.

One specialist put it bluntly:

"I don't think you can get such a strong model this fast right after

Fable if all you're doing is distillation," one of the experts told TechCrunch.

The point of the remark is that a "suspiciously good" result is not evidence in itself: the expert attributes Kimi K3's speed and quality not to borrowing but to the team's work. The phrase "right after Fable" is key here: the expert is pointing to too short a time gap for pure distillation, in his view, to have produced such a jump in quality.

What model distillation is

Distillation is a training method that involves two models: a strong "teacher" and a trainable "student." The developer runs queries through the powerful "teacher" model, collects its answers, and trains its own "student" model on this data, replicating behavior without access to the original weights or training data.

The method is cheaper and faster than training from scratch, but it has a natural ceiling: the student on average does not surpass the teacher and inherits its weak points and error patterns. So if Kimi K3 were merely a distillate of Fable, it would hardly look as convincing as it has surprised observers.

The distillation dispute also has a legal dimension. The terms of use of leading AI labs, including Anthropic, generally prohibit training competing systems on the outputs of their models. So an accusation of distillation is not just a technical question but also a potential claim of a rules violation. However, there is no public evidence yet that Kimi K3 was trained on Fable's answers.

Where did the suspicions come from?

Suspicions about Kimi K3 arose from a combination of speed and strength: the model appeared shortly after Fable's release and immediately showed a high level of performance. In the industry, accusations of distilling other companies' top models come up regularly — especially against fast releases that are catching up with Western leaders.

Such disputes are nothing new for the industry. When a team catching up with the leaders releases a strong model, part of the community almost automatically assumes it borrowed from bigger players — as happened with a number of previous open and Chinese releases. Kimi K3 fell into the same logic: if the result exceeds expectations, then there must be a "teacher" somewhere. The experts polled by TechCrunch suggest not rushing to that conclusion.

Proving distillation after the fact is difficult: characteristic "fingerprints" of the teacher model sometimes surface in the student's answers, but there is no reliable, universally accepted test for this. The experts polled by TechCrunch essentially separate two different claims: that Kimi K3 is indeed strong, and that the reason is specifically copying Fable. They do not dispute the former; they consider the latter unproven and unlikely.

What this means

The dispute around Kimi K3 shows how hard it is to prove the origin of a strong model from the outside. On the surface, "too fast" progress is easy to mistake for copying, but no one has produced direct evidence of Fable distillation, and the experts polled on July 23, 2026, lean toward the theory of the team's independent work. For the market, this is a reminder: a strong result by itself is not yet evidence.