Codewhale Enterprise
Own the intelligence your team runs on
Your team's work already runs through Codewhale. Enterprise turns that work into evals, training data and models you own. It is in development and open to design partners.
What you get
Every model is tested on your own held-out tasks before it replaces anything.
Evals that define done
An eval suite built from your team's real work, with a held-out set that decides whether a model is good enough.
Training data from your work
Data drawn from work your team already accepted, with sensitive details removed.
A model you own
A fine-tuned open-weight model you control, compared with your current model on quality, cost and speed.
Environments that keep improving it
RL environments built from your own tasks, so your model keeps improving on the work that matters to you.
Available today and planned
The open-source agent is available now. Enterprise is in development, and no customer model has been trained yet.
Available today
An agent for any model
Hosted providers, any OpenAI-compatible endpoint, and models you serve yourself with vLLM, SGLang or Ollama.
Saved sessions with receipts
Every session keeps its tool calls, results and approvals, so your team's work is on record.
Planned for Enterprise
Managed eval and training environments
Environments that run your evals and training without depending on a laptop.
Qualification on every new base model
When a better open model ships, it is tested on your holdout and promoted only if it passes.
Routing between rented and owned models
Each workflow goes to the model, rented or owned, with the best accepted result per cost.
Private deployment
Run the agent and Enterprise where your data must stay.
Your data stays where you choose
Your sessions, datasets and models stay in places you choose. Training runs only on compute you approve.
One Rust engine with approvals
The same engine runs your workflows and Enterprise. Ask, Auto-Review and Full Access permissions and receipts apply to both.
Design partners
Bring one workflow your team repeats. You get a baseline of your current model on your own holdout, then a test of an owned model against it. Email us to start.