Le Chaton Fat: How a Fake Mistral Model Fooled the AI Internet

The Model That Didn't Exist

On June 11, 2026, a parody benchmark chart began circulating on X. It showed a plump white cat — unmistakably a riff on Mistral's "Le Chat" chatbot — topping a leaderboard with stats no real model has ever claimed: 100 trillion parameters, a score that put it thirty points above Fable 5 on something called "VoltaireBench," and a safety profile described simply as "très bon."

The chart was obviously a joke. Except — it wasn't entirely obvious. And that's what made it interesting.

By June 12, the replies had fractured into three camps: people who knew it was satire, people who were genuinely unsure, and people who were 100% certain it was real and furious they couldn't access it because of EU export controls. Researchers at named labs were quietly DMing each other. "Is Le Chaton Fat real?"

It was not. But the story of why so many smart people entertained the question for even a moment is worth unpacking.

How It Started: Le Chat Becomes Le Chaton Fat

The joke builds on a pun that works better in French. Mistral's consumer chatbot product is called Le Chat — "the cat." The parody model name Le Chaton Fat translates roughly to "the fat kitten" or "the fat little cat," depending on your accent. The visual gag — a round, self-satisfied white cat in the style of AI company mascots — was charming enough to screenshot and share.

The original parody chart kept the specs deliberately absurd: 100 trillion parameters, a benchmark score that didn't map to anything real, and a context window labeled "∞ (for croissants)." It was clearly satire aimed at AI benchmark inflation culture.

But the internet, as the internet does, started iterating.

The Rumour Stack

Within 48 hours, the joke had accumulated a full fake spec sheet that sounded increasingly plausible:

• 30-trillion-parameter Mixture-of-Experts architecture with 256 active experts
• 1-million-token context window
• Multimodal and multilingual by default
• "Napoleon class" — a new fictional model tier coined by @fabianstelzer
• Outperforms Fable 5 on every benchmark (including the real ones, per parody posts)
• Safety constraint: "for safety reasons only outputs French language code"

That last detail — the French-only code output — was the tell for most readers. But the first four specs? They sit comfortably inside the range of what real frontier models announce. A 30T MoE with 256 experts and a 1M context window is plausible. That's the problem.

@AlexanderKnigge posted a straight-faced "confirmation" that Mistral had officially announced the release, with the full spec list. No quote-tweet, no 🧵 label, no /s. It racked up thousands of impressions before the replies caught up.

The CEO Adds Fuel

The moment that genuinely muddied the water arrived when Mistral co-founder and CEO Arthur Mensch replied to a thread about it:

"It's actually le gros chaton"

Le gros chaton is French for "the fat cat" — grammatically a correction of the parody name, tonally a joke. But 280 characters stripped of vocal inflection can read many ways. Was he correcting the name because the model exists and the name was wrong? Was he just playing along?

French media [@Mamading Ceesay] couldn't find the model on Mistral's website and concluded it was a faux modèle — a fake. Cody Blakeney (@code_star), a researcher at Arcee AI who'd been following the thread, tweeted the question directly: "Can someone tell me if Le Chaton Fat is real or an amazingly elaborate joke?"

He later apologized: "I realize my mistake. Obviously it's real, and I don't have access because of EU export controls."

Then, a few hours later: "I apologize for asking such a silly question." (He had confirmed it was fake.)

Why It Hit Different in June 2026

The Le Chaton Fat meme landed at a specific cultural moment that made it stickier than it would have been at any other time:

1. The Fable 5 export ban created a sense of chaos. Anthropic had just been forced to suspend Fable 5 globally under a U.S. Commerce Department directive (read our full Fable 5 ban breakdown). The AI community was already primed to believe that surprise announcements were possible and that access to top models could change overnight.

2. Mistral has a history of unannounced drops. The French lab has shipped models with minimal pre-announcement before. The "this sounds like something they'd do" instinct wasn't unreasonable.

3. EU/US tech rivalry was already the week's narrative. @eurofounder posted, apparently straight-faced: "While Anthropic is fighting with the US government, Europe quietly built the most powerful AI in the world... We must now ban access to Mistral for all Americans." The joke played perfectly into the geopolitical frame everyone was already using.

4. Benchmark scores are mostly vibes anyway. The community has developed such deep skepticism toward leaderboard numbers that "this model scored X on VoltaireBench" barely registers as a claim requiring verification. Half the joke is that VoltaireBench is fake — but half the audience wasn't sure it wasn't real either.

What Mistral Actually Released

While Le Chaton Fat was dominating timelines, Mistral's real product lineup stayed quiet. Their actual June 2026 releases are:

• Mistral Medium 3.5 — their latest mid-range model for enterprise workloads
• Mistral Small 4 — optimized for cost-efficient inference at scale

Neither is "Napoleon class." Neither has 30 trillion parameters. Both are available through the Mistral API and Le Chat. You do not need to be outside the EU to access them. They do not output exclusively French code.

The Satire That Was Actually Serious

Underneath the joke is a real critique that the AI community absorbed and retweeted because it stings:

Benchmarks mean almost nothing. A model can top a leaderboard most people can't name on tasks nobody ships to production, and the announcement gets treated as a major geopolitical event. Le Chaton Fat — a fictional cat — almost got the same treatment.

"Confirmation" is a broken concept in AI news. A CEO reply that is clearly a joke can read as product confirmation. A parody spec sheet, if precise enough, becomes indistinguishable from a real announcement. The gap between satire and press release has collapsed.

Export controls created paranoia. The EU/US split over model access, made suddenly concrete by the Fable 5 ban, primed people to believe that a powerful model might exist somewhere they just can't see. Cody Blakeney's instinct — "I don't have access because of EU export controls" — was a completely rational explanation in a world where that is now a real thing that happens.

How Le Chaton Fat Compares to Other AI Hoaxes

Le Chaton Fat is not the first fake model to briefly fool AI Twitter, and it won't be the last — but it's worth placing next to its predecessors to see why it spread faster and lasted longer than most.

What set Le Chaton Fat apart is that it never tried to be fully convincing — it was satire from the first post — yet it survived scrutiny anyway because the ambient conditions (export bans, unannounced drops, benchmark fatigue) did the work that a more careful fake would have needed forged screenshots to do. That's arguably a more interesting finding than the hoax itself: you no longer need convincing fabrication to create real confusion. You just need a plausible number dropped into an already-anxious news cycle.

What This Reveals About AI Benchmark Culture

The Le Chaton Fat episode is a useful stress test for how the AI industry actually evaluates claims in 2026, and it exposes a few uncomfortable patterns:

Benchmark names carry more authority than benchmark methodology. "VoltaireBench" sounds like it could be real — a French-flavored eval to match a French lab — and that surface plausibility was enough to get people repeating the score without asking what VoltaireBench measures, who built it, or where the eval set lives. Compare this to how real benchmarks like MMLU, SWE-bench, or ARC-AGI earn scrutiny: people ask about contamination, task design, and reproducibility. VoltaireBench got none of that scrutiny because nobody paused to check it existed.

Parameter counts function as a headline number, not a technical spec. "100 trillion parameters" and later "30 trillion, 256 experts" register as impressive before anyone asks whether that's dense or MoE-active, what the inference cost would be, or whether any lab has published training infrastructure at that scale. Real frontier labs are typically cagey about exact parameter counts precisely because the number alone tells you almost nothing about capability — but the hoax worked because audiences still treat it as the single most important spec.

Ambiguity from public figures gets over-interpreted as confirmation. Arthur Mensch's "It's actually le gros chaton" is a joke, grammatically correcting a French pun. Read charitably, it is obviously not a product announcement. Read anxiously — by people already primed by the Fable 5 export ban to expect surprise moves — it reads as a wink-and-nod confirmation. This is the same dynamic that fuels stock-market rumor mills: an executive's silence or deflection gets treated as signal, when it's often just someone enjoying a joke.

What Comes Next

Le Chaton Fat will not be released. There are no weights, no API, no product page. The plump white cat exists only in screenshots and in the part of your brain that briefly entertained the possibility.

But the conditions that made the hoax work — benchmark inflation, opaque release cycles, geopolitical tension around model access, and a community trained to believe anything could drop at any moment — those are not going away.

The next Le Chaton Fat will be slightly more believable. And the one after that more believable still.

In the meantime: Mistral's real models are available at mistral.ai. They are good. They are not 30 trillion parameters. They do not exclusively output French code.

Related Reading

• The Fable 5 U.S. Export Ban — Full Breakdown
• When Will Fable 5 Be Available Again?
• GLM-5.2: China's Real Response to the Fable 5 Ban
• Kimi K2.7: Moonshot AI's Open Coding Model