Blog

Product · · 4 min read

Why Ineffable runs on open-weight models

We chose open-weight language models for every specialist. Here is why that matters for data control, cost and private deployment.

When we started building specialists, the obvious choice was to call the biggest closed model available and move on. We chose open-weight models instead. Here is the reasoning.

Your data stays out of training

Companies are being asked to put tariff files, payroll data, medical claims and acquisition financials into AI tools. The first question every security team asks is where that data goes and whether it will be used to train a model.

With open-weight models, the weights are fixed and published. We run inference, return the answer and keep nothing in the model. We do not train on your workspace content, and we do not use inference providers that do.

Private deployment is real, not a promise

Because the weights are available, an enterprise customer can run its specialists on dedicated GPUs, isolated from everyone else. For regulated industries, that turns "trust us" into an architecture you can review.

The model is not the moat

The difference between a generic chatbot and a useful specialist is not the last few points on a benchmark. It is the playbook: the six steps a senior practitioner follows, the expert checks at each step, the sign-off points and the shape of the final deliverable. That knowledge lives in Ineffable, not in the model. Open models are now strong enough to execute it well.

Cost that scales with your team

Specialist work involves long documents and many back-and-forth turns. Efficient open models keep the cost per session low, which is what makes per-seat pricing for a whole team possible.

What we do not use

We deliberately avoid models whose development or hosting raises data-residency questions for our customers. We would rather give up a benchmark point than make a security review harder.

We will keep evaluating new open-weight models as they are released, and switch when one does the work better. You can read more about how we handle data on our security page.