The name carries a spirit forward.

Kapllan Labs takes its name from Kapllan Hasani — a name rooted in Kosovo and carried forward as a symbol of courage, duty and protection.

For us, Kapllan represents more than a name. It represents the values of the knight: strength with responsibility, intelligence with discipline, and power used in service of something larger than oneself.

Kapllan Labs was founded with that spirit — to build AI capability that is not distant, borrowed or disconnected from the people it affects, but developed close to our languages, institutions, industries and real-world problems.

The lab carries this origin into the future: transforming inherited values of courage, protection and service into a new kind of Nordic AI capability.

Kapllan Labs exists to build enduring AI capability in the Nordics.

We are not building toward a single app or a single model. We are building the knowledge, methods, technical competence and systems understanding required to develop advanced AI capability over time.

We believe critical AI competence should not only be rented from external systems. It must also be developed locally — close to our languages, institutions, industries, values and real-world problems.

Kapllan Labs is a lab where research, engineering and practical application come together. The systems, models, tools and products that emerge are the outcome of that process working.

We are building the capability, not just the product.

K—1 is learning to read.

A 1.5-billion-parameter dense model, training on a carefully filtered Nordic-leaning corpus. It is the first rung of the K-series ladder — small enough to iterate on, serious enough to teach the lab what its stack does under load. Quality gate at day 14 decides whether K—2 starts.

Parameters

1.5B

Dense decoder-only. GQA + QK-norm.

Tokens seen

60B

Filtered web, code, books, papers.

Context

Extended to 128K in mid-train phase.

Compute

8× H100

Stockholm DC-01 · BF16 · FP32 master.

Wallclock

14d

Day 9 of ≈14. ETA 4d 11h.

Gate

d14

Reproducibility · loss · eval suite.

Now training

K—1reader

First rung. Small, careful, reproducible.

1.5B · 60B tokens
8 × H100 · 2 weeks

Queued · Q3

K—2generalist

Awaits K—1 quality gate.

8B · 2.5T tokens
32 nodes · 6 weeks

Planned · Q1 '27

K—3frontier dense

First model intended for outside scrutiny.

32B · 8T tokens
128 nodes · 10 weeks

Planned · Q4 '27

K—4frontier MoE

120B total, 15B active.

12T tokens · 64 experts
256+ nodes · 14 weeks

Read the brief. Or come to the training floor.

Live training floor →Research notes Llana 3

K—1

A model that reads carefully, and asks before guessing.

The name carries a spirit forward.

Kapllan Labs exists to build enduring AI capability in the Nordics.

Three envisioned products, one word.

Designed down to the letter.

Pllan

One endpoint, every model. Predictable costs to two decimals.

Llana

A model that thinks patiently, in public.

Ofe

A voice that doesn't slip on the third hour.

K—1 is learning to read.

Read the brief. Or come to the training floor.