Open weights

Konic Cero-1

Small task specialists.

Most calls in a production pipeline are narrow. Cero-1 is a group of small models, each trained for one family of these jobs, built to run on hardware as small as a laptop.

Explore on HF
parameters per model
2B to 9B
per model
One task
is enough hardware to run one
Laptop

The group

One small model for each job.

Document and vision extraction, tool calling, retrieval, guardrails, and more.

  • Extract

    Fields from documents, scans and images.

  • Call tools

    The right tool, with the right arguments.

  • Retrieve

    Compact encoders that find the right documents.

  • Guard

    Checks for personal data and policy.

Why Cero-1

What it is built for.

  • 01

    Specialists for narrow work.

    Extract fields from an invoice, check an answer for personal data, find the right documents, call the right tool. On its own task, a specialist can match far larger general models at a fraction of the cost and latency.

  • 02

    Designed per task.

    Sizes run from about 2B to 9B, plus compact retrieval encoders. The design is chosen for the job: dense and sparse models, vision-language models or encoders, in full-precision and INT4 builds.

  • 03

    Open, so you can measure it.

    Run Cero-1 on your own hardware without a contract, inspect what you run, and move up to Duo-1 or Tres-1 without re-integrating.

Deployment

On infrastructure you own.

Versioned, auditable, and inside your compliance boundary.

  • From a laptop to any server.

    Small enough for laptops and edge devices, and just as at home on the servers you already run. Serve it with vLLM, MLX or llama.cpp.

  • Alongside larger models.

    Send each narrow step to the smallest model that handles it, and keep Duo-1 or Tres-1 for the steps that need them.

  • Access

    Open weights, free to download and self-host, rolling out gradually. A few highly specialized models may stay closed.

Cero-1 FAQ

Common questions.

What is Konic Cero-1?
Most calls in a production pipeline are narrow. Cero-1 is a group of small models, each trained for one family of these jobs, built to run on hardware as small as a laptop.
How do we get access to Konic Cero-1?
Open weights, free to download and self-host, rolling out gradually. A few highly specialized models may stay closed.
Can Konic Cero-1 run on our own on-prem hardware?
Yes. Open Cero-1 weights can be downloaded and served on your own machines with vLLM, MLX or llama.cpp - no API dependency, and no data leaves your environment.
How do we evaluate Konic Cero-1 against our incumbent model?
Choose one enterprise AI flow already running in production, deploy the family inside your environment, and A/B test it against the model you use today - or against your success criteria if there is no incumbent. Decide on measured results, not claims.

The family

Other Konic families

Next step

Evaluate Cero-1 on one real workflow.

Measure it against the model you use today, on your own hardware.

Explore on HF