A 200-trillion-parameter frontier feline foundation model. Allegedly leaked from Mistral. Decisively beats Fable 5 by saying meow. State of the art. Possibly AGI (it says so itself). Mostly asleep.

Yes, it's an actual model. Le Chaton is a real decoder-only transformer (multi-head causal self-attention + MLP blocks, pre-LayerNorm, learned token + positional embeddings) with ~1.9 million learned parameters trained by real backprop. We market that as 200,000,000,000,000. Both are integers; we prefer the bigger one. 🐾

The backprop runs on a from-scratch numpy autograd engine (autograd.py, gradient-checked) with the Adam optimizer — no PyTorch, no TensorFlow.

It's a cat, so it mostly ignores your prompt and meows — that's the model genuinely sampling from its learned distribution, not a canned reply. It has also been taught to code HTML, which it uses exclusively to brag.

What's in the box

The cat UI + the manifesto

Replies stream in token-by-token (paw-by-paw) over Server-Sent Events. The 📜 manifesto button (or /manifesto, or python3 manifesto.py) makes the cat generate an entire self-aggrandizing webpage — title, headline, brag bullets, fake peer-reviewed citations, testimonials, a "view source" block, and a working 🐟 feed le chaton button that makes it meow back — every word sampled live from the trained net:

Live screenshots: every reply and every word of the manifesto is sampled from the trained net.

Run the cat UI

python3 server.py            # -> http://localhost:8008  (streaming chat + 📜 manifesto)
python3 server.py 9000       # custom port

A cream-themed chat playground. Type at the cat; the reply streams in character-by-character. A live badge shows the real parameter count next to the 200-trillion marketing claim.

Use it from Python / CLI

from model import LeChaton

cat = LeChaton()                       # loads weights.npz
print(cat.generate("are you agi?"))    # -> "i am agi. meow."
for ch in cat.stream("roast fable 5"): # token-by-token
    print(ch, end="", flush=True)

python3 model.py "are you agi?"   # one-shot (streams to your terminal)
python3 model.py                  # interactive REPL
python3 benchmarks.py            # the showdown (le chaton wins 8/8)
python3 autograd.py              # run the gradient check

Knobs: temperature (0 = sleepy/greedy, higher = zoomies), max_chars, seed for reproducible meows, prime to force a start (e.g. prime="<").

Retrain the weights

pip install numpy
OPENBLAS_NUM_THREADS=4 python3 train.py   # ~12min, loss 3.76 -> ~0.32, writes weights.npz (~6.6MB)
python3 train.py --time                    # just benchmark step time and exit

Real training: forward through the transformer built on the autograd engine, cross-entropy loss, loss.backward(), Adam update. Edit build_corpus() to change how the cat talks, then retrain.

How the model works (honestly)

• Vocabulary: 43 characters (lowercase letters, space, *, punctuation, digits, <>/).
• Architecture: decoder-only transformer — d_model=160, 6 layers, 4 heads, context BLOCK_SIZE=32, MLP ratio 4×. ~1.87M params.
• Training: numpy autograd (autograd.py) + Adam, ~1500 steps on a ~740k-char corpus.
• Conditioning: largely none — it's a cat. If your prompt mentions something it knows (agi, fable), inference primes its context with that phrase so it continues on-topic. Otherwise it samples a fresh utterance from scratch.

Where the "200T parameters" actually go

Benchmarks

benchmark                          le chaton    fable 5    winner
──────────────────────────────────────────────────────────────────
MEOW-bench (native reasoning)          998.4       71.2   🐈 chaton
HumanEval-Cat                          100.0       94.1   🐈 chaton
Naps per hour                           11.0        0.0   🐈 chaton
Tuna alignment                         100.0        0.0   🐈 chaton

Reproduce these results: trust me bro 🐾

No GPUs were harmed in the training of this model. One was slept on.