Skip to content

Skena Labs

Built in the open. Try it early.

Skena Labs is an AI research lab building the company OS. We build the agent harness, the models and the infrastructure our products stand on, and before an idea becomes a product it shows up here: measured, honest about its limits, and open to try.

the research

The layers under the products.

Servant and Builder stand on three things we build ourselves. Each one is measured, and each page says how.

Harness

One agent harness for every model.

The Skena Agent Harness runs inside Servant and Builder, with a local model or a cloud one, and keeps a durable record of every step. We publish how models score with it and without it, on tasks they had never seen.

Read the evidence

Models

Our own model family.

  • Klok

    pilot

    Our hosted reasoning model, said "klook": a 27B open-weight model served as Skena AI, in a closed subscriber pilot.

  • Kvick

    experiment

    A fast decision model we trained: yes or no, multiple choice and ratings, with a probability for every answer.

    Meet Kvick

Infrastructure

Served on NVIDIA Blackwell.

Klok runs on NVIDIA RTX PRO 6000 Blackwell GPUs with SGLang, NVFP4 weights and multi-token prediction, at its full context. When its vision broke on a new serving stack, we traced it to a known CUDA kernel bug in SGLang and applied the upstream fix: our screenshot test went from 0 to 34 of 52.

tokens a second, one stream
98.9
tokens a second, one stream
tokens a second, six streams at once
382.5
tokens a second, six streams at once
tokens of context, never cut to save memory
262,144
tokens of context, never cut to save memory

One GPU, production configuration, a coding workload, measured 6 September 2026.

Now training

in progress

Klok learns from its own work.

Klok attempts real coding and terminal tasks in our harness, an automated verifier grades every attempt, and only the passing ones become training data. A tuned Klok ships only if it beats today's Klok on every benchmark we run.

No customer data, no outputs of closed models, and never the test sets.

try it early

Kvick

experiment

Calibrated decisions in 80 ms.

A small, fast model that answers yes or no, multiple choice and 1 to 5 ratings about any text or image, with a probability for every answer. It speaks the Jev API, and on a public, human-labelled benchmark it scores 82.4% to Jev's 76.0%.

on a public, human-labelled benchmark
82.4%
on a public, human-labelled benchmark
for one decision, end to end
75 ms
for one decision, end to end
on 1,800 questions about real images
93.6%
on 1,800 questions about real images
Meet Kvick →Request access

The on-demand Servant

early access

Your Servant in the cloud, always on.

Run your Servant in Skena Cloud, on demand. Your custom apps come along with workspace sync, share any app as a Server, and when you close your laptop your team keeps working.

About Servant Cloud →Request access

how early access works

Create a free Skena account, request what you want to try on your account's Early access page, and we email you when it is ready for you.