01
Frontier models send your data out.
The strongest models are closed. You can only reach them through the provider’s API, so your data goes to the provider.
A private AI workspace that keeps getting better at your company’s work.
It runs inside your environment, on the data you choose. It learns from your data and from its own work, and keeps only the changes that score better.
Now taking on our first design partners.
ResearchA 9B open model, trained for the task, matched GPT-5.5 medium on our SQL test: 115 of 220.
Held-out scoreno change
22%
Round00
Versionv01
DecisionBaseline
Data sent out0 B
Illustration of the loop. Not customer data.
The problem
Why AI isn’t doing your company’s work yet.
Three options. Each one has a problem.
02
Big open models are hard to run.
Running them yourself is expensive and needs a team that can manage the infrastructure at that scale. Renting them from an inference provider sends your data out again.
03
Small to mid-size models are not good enough off the shelf.
They are affordable and run in your environment. Out of the box they are weaker, and they don’t know your data or how your team works.
Nablo makes small to mid-size models good at your work.
It trains the model and its harness on your data and your work, in your environment, and keeps improving them. It can match a frontier model on your work. You keep the trained model.
A 9B open model, trained for the task, matched GPT-5.5 medium on our SQL test: 115 of 220.
Self-improving
It learns from your data, then from its own work.
- 01
Connect the data you choose.
Your documents and databases stay where they are.
- 02
Nablo turns it into practice tasks.
Questions drawn from your data, each with a verified answer.
- 03
It improves itself on those tasks.
It tunes its instructions, tools and memory, and trains the model itself when that helps more. A change stays only if it scores better on tasks it has not practiced on.
- 04
It keeps learning from real work.
Every run leaves a trace of what worked and what failed. Nablo learns from those traces and improves again, round after round.
Step 1 of 4
Selected sources only. Nothing is copied out.
The loop runs on its own. Nobody labels data or writes prompts by hand. Each new version waits for your review.
Research
We publish our experiments, including what did not work.
Prompt optimization · Sep 2026
2 → 6of 9 unseen tasks
Before
After
Nablo generated the practice tasks from a demo customer’s database. The winning change was an eight-line instruction edit, on one local open model. No data left the machine.
Read the experiment →Small-model distillation · Aug 2026
115of 220 tasks
Qwen3.5-9B matched GPT-5.5 medium’s aggregate score on Nablo’s frozen 220-task SQL-agent evaluation. Selected checkpoint, not a holdout estimate.
Read the case study →Internal research. Not customer results.
Privacy and control
Your environment is the boundary.
It runs where your data lives. Nablo runs in your own environment: your servers, your cloud, wherever you choose. It reads only the sources you select, and it never writes to them.
You choose where the model runs. By default, an open-weight model runs in your environment.
Nothing else leaves. There is no telemetry and no fallback to an outside AI service. Anything outside your environment works only through a connection you set up.
Start a conversation
Start with one piece of work.
Tell us the work and the data it needs. We will show you how Nablo would learn it inside your environment.
isaac@nablo.io