Why experts doubt Kimi K3 is just distilled Fable

White House science advisor Michael Kratsios publicly accused Moonshot AI of building Kimi K3, the largest open-weight model available, by copying Anthropic's Fable.

He also claimed the Chinese lab trained it using smuggled Nvidia Grace Blackwell 300 (GB300) chips and Thai server farms. Treasury Secretary Scott Bessent quickly chimed in, claiming officials are finding digital watermarks of U.S. frontier models inside Chinese systems, fueling talk of a total ban on Chinese open-weight models.

There is just one problem: AI researchers took one look at the timeline and called BS.

The math behind a two-week distillation doesn't work

Braden Hancock, co-founder of Snorkel AI and researcher at the Laude Institute, pointed out the obvious calendar conflict: Anthropic only released Fable on July 1st. Pulling off full capability distillation, data cleaning, architecture pre-training, and evaluation for a model this massive in just two weeks is practically impossible.

Nathan Lambert, an AI researcher at the Allen Institute for AI (AI2), echoed that skepticism on a recent podcast. He noted that as Chinese labs close in on the frontier, simple supervised fine-tuning (SFT) distillation offers sharply diminishing returns. SFT mostly teaches a model its "manners"—which explains why poorly fine-tuned models occasionally blurt out that they are Claude—but it doesn't build real reasoning muscles.

Distilling actual Fable-level frontier capabilities would require massive reinforcement learning (RL), where the teacher model grades tens of millions of agent interactions.

Running that over an API is not only wildly expensive, Lambert explained, but the severe API latency would choke training throughput, likely resulting in zero performance uplift. While Anthropic previously flagged millions of suspicious capability extraction queries from Moonshot, DeepSeek, and MiniMax earlier this year, distillation is standard across the board. Elon Musk even admitted under oath that SpaceXAI distilled OpenAI outputs to train Grok.

A 2.8T beast that goes far beyond API scraping

The claim that Kimi K3 is merely a distilled clone also collapses once you look under the hood.

The weights uploaded to Hugging Face reveal a monstrous 2.8-trillion parameter Mixture of Experts (MoE) model with 104 billion active parameters. It runs on custom-designed Kimiko delta attention, attention residuals, and Moonshot's native Moon VIC vision encoder.

The raw model weighs in at a staggering 1.56 terabytes, requiring serious enterprise GPU clusters just to host. The open-source community has already stepped in, with Unsloth publishing a 1-bit GGUF quantization that shaves the footprint down to 594 GB, though it still demands heavy compute.

Hancock emphasized that U.S. commentators routinely underestimate Chinese technical teams, noting Moonshot's founders hold PhDs from elite institutions like Carnegie Mellon University.

As for the hardware allegations, Sam Bresnick of Georgetown's CSET acknowledged that chip smuggling exists, pointing to the federal indictment of Supermicro's founder in May. But while Bresnick advocates for strict "know your customer" rules for offshore data centers, Biden-era 2024 proposals have stalled under Donald Trump.

Why it matters

We see Kratsios's broadside less as an airtight technical indictment and more as political anxiety over open-weight parity.

Kimi K3 is an undeniable technical heavyweight. While a 1.56 TB checkpoint is entirely out of reach for consumer workstations, making the raw weights freely accessible on Hugging Face gives researchers and enterprises a frontier-grade foundation that rivals top proprietary systems.

Dismissing genuine MoE architecture breakthroughs, custom attention mechanisms, and RL infrastructure as a quick two-week copycat job simply ignores how post-training works today. SFT is cheap, but scaling frontier reinforcement learning requires massive engineering grit.

In our view, attempts to weaponize distillation allegations to justify banning open-weight releases will do little to slow adoption. When export curbs fail to prevent the rollout of 2.8T models, political friction is inevitable. But in the open-source arena, real architectural execution will always outlast regulatory finger-pointing.