Post on X by Brian Armstrong, CEO of Coinbase: We were able to fine tune an open weight model (Qwen3.5-9B) on proprietary internal data at Coinbase, and it out performed frontier models, at much lower cost and speed, for a specialized task (fraud prevention in this case). I think we'll see more of this pattern inside companies. The frontier models are amazing generalists. But increasingly you'll see specialist models out perform on more narrow tasks, either working on teams, or by themselves. The attached chart shows the fine-tuned model improving fraud detection over each comparison model, including an Opus 4.5 decision agent.
Brian Armstrong, CEO of Coinbase, on X ↗
Active learning and the data flywheel

Specialist models
can beat generalists.

Coinbase fine-tuned Qwen3.5-9B, an open-weight model, on its own data. For fraud prevention, it outperformed frontier models at much lower cost.

Evals and active learning

Short on engineering, but want to train your own model?
We can help.

RadianVector focuses on optimizing inference for open models. We also build the products around your training loop: evals, datasets and versioning.

You train new versions of your model, and we evaluate each one. We tell you how good it is and how it compares with the model it started from, before you customized it.

DATA FLYWHEEL You traina new versionWe evaluate iton your tasksWe showthe weak spotsYou add datafor those parts youRadianVector
  • Where the model is goodThe tasks the new version gets right.
  • Where it is badThe tasks it gets wrong, with examples.
  • Which part needs more workWhere more training data will help most in the next round.
  • How it comparesAgainst the model it started from, and other models you could use instead.

Products around your training loop

  • Pairwise evalsTwo versions on the same tasks, side by side.
  • Stability checksThe same request repeated, to see how often the answer changes.
  • Failure miningThe cases your model gets wrong, found and grouped.
  • Long vs short contextHow quality holds as the prompt grows.
  • Eval datasetsBuilt from your tasks, with checked answers.
  • Held-out datasetsKept out of training, to score each version on data it has not seen.
  • Model versioningEvery version, with the data that went into it.
  • Results versioningEvery score tied to the model version and dataset behind it.
Start a conversation

Training your own model?
Let’s measure it.