Vestra Deepfield Open Deepfield Studio

Vestra Deepfield.

Vestra’s AI research and product team, named after the Hubble Deep Field: the photograph that found thousands of unseen galaxies by looking further than anyone had. We build Vestra’s AI systems on hardware we own and run, in the UK.

Open Deepfield Studio See the models

Deepfield Studio

The developer studio for Vestra’s models, live at platform.vestra.systems. VestraAI is the customer-facing product.

Playground
Model picker, streaming replies and a short onboarding. Free
Presets
Balanced, Precise, Creative and Code for everyone. Custom presets on Plus and above
Compare and saved prompts
Two models side by side, and prompts you keep. Plus and above
View code
curl, Python, Node.js (OpenAI SDK) and JSON. API keys are never written into the snippets.
Structured output
json_object and json_schema in the Playground and the /v1 API, with schemas up to 64 KiB, on llama-server models.
Models, usage and docs
A models page, usage and cost dashboards, and the API documentation.

64,000

characters a message on Free and Plus. Prompts on these plans are capped at 64K tokens.

384,000

characters a message on Max 10x and Max 20x with long context on. Input past 64K tokens is charged at 2×.

Plans, per month, in pounds
FeatureFreePlusMax 10xMax 20x
Price£0£10£50£100
Playground and built-in presetsYesYesYesYes
Compare, custom presets, saved promptsNoYesYesYes
Characters a message64,00064,000384,000384,000
Long context, over 64K tokensNoNoYesYes

Long context is for Max only, while it is switched on; API keys paying with credits are the documented exception. See the docs for limits and prices.

Models and infrastructure.

Inference is hosted in the UK, on a GPU box Vestra owns and runs, served with llama-server behind platform.vestra.systems. It is not a rented public-cloud model API.

  1. vestra

    Vestra

    The general model.

    15pinput, per million tokens

  2. robot-1.1-coding

    Vestra Robot 1.1

    For coding, with vision. Finetuned Qwen 3.8.

    15pinput, per million tokens

  3. cyber-mini

    Cyber Mini

    By grant only.

    20pinput, per million tokens

  4. cyber

    Cyber

    By grant only.

    25pinput, per million tokens

Cached input
25% of the input rate, for the part of a prompt the server reuses from its cache.
Output
£1 to £2 per million tokens, set by how busy the box is when a request starts; higher on the Cyber models.
Hosting
UK-hosted, on hardware Vestra owns and runs.

Updates.

What has shipped, newest first.

  1. Deepfield Studio is live.

    The Playground, Compare, presets and saved prompts, View code, structured output, models, usage and docs, at platform.vestra.systems.