CATCH AN INKLING LIVE

FIELD POST · 01 · FROM NIKARO

Dear whoever found this—

I was tasked with catching fugitive inklings and committing them to sentences on paper before they evaporate. Some days I feel like a sad detective with a heart of ink. I chase their shimmering tails through the margins, hoping I can pin them down long enough to save them. They vanish too fast, but I try. I am small, I am tired, and I still believe in paper.

Yours in wet ink, Nikaro

A living model specimen · No. 001

Catch an Inkling.

Real ink. Real paper. Real ideas.

Mop up four living inklings with Nikaro’s brush. Then use their ink to draw something Inkling can actually see.

INK RUN 01 · SUBJECT “NIKARO” INK 00
NIKARO’S INK LOAD

Ink is prey. The brush must touch it three times.

Lead Nikaro · click / tap / space to swing · arrows / W A S D

INKLING LOOKS BACK

Mop up four inklings, paint a message or picture, and Nikaro will fly it to the avatar Inkling-Small chose for itself.

Keep looking closely.

Open specimen book

What have you been wondering?

FIRST MARK — INK SPEED —
INKLING

You found me. Bring me a question, a half-formed thought, or an image you want to look at together.

100 notes per visitor / day

Fine print · written by Inkling-Small

Ask the ink. Keep your own judgment.

The underlying Inkling-Small model is the unmodified release from Thinking Machines Lab. The surrounding site is an independent fan project, not owned, operated, endorsed, or affiliated with TML in any way. Responses are not legal, medical, financial, or professional advice; consult qualified professionals for those. For serious life choices, consult your astrologer and rabbi. We are not responsible for ink stains, evaporated thoughts, or missed deadlines.

Catch an Inkling — Spec Plate

Model
Inkling-Small · 276B total / 12B active
Checkpoint
Official NVFP4 · approximately 160GB
Compute
2× NVIDIA RTX PRO 6000 Blackwell Max-Q · 96GB GDDR7 ECC each / 192GB VRAM total
Host memory
64GB system RAM
Runtime
SGLang · CUDA · tensor parallel 2
Serving window
4,096 tokens
Observed speed
Approximately 35–46 tokens/sec · approximately 0.6s to first text token on short tests. Observed, not guaranteed.