New playbook: when your own AI model pays. Read
Mercor (expert data) reported this on 15 Oct 2026 in company self-report (source). It belongs to the learning loop layer.
NVIDIA Nemotron 3.5 Super post-trained on a few hundred expert tasks per domain (legal, software engineering, accounting) from Mercor's APEX work environments.
On held-out tasks, versus the base model: legal rubric score +6.7 pp, software full-pass +8.2 pp (26.6% to 34.7%), accounting rubric score +7.3 pp.
The number covers APEX benchmarks; measured against the base model; a comparison with frontier models is not reported.
“With only a few hundred training tasks per domain, that's a meaningful gain on work the model had never seen,” Mercor said.
Reviewed corrections are kept as test cases and later used to train models.
96%
lower estimated serving cost for one agent
A daily continual-learning loop for Sidekick's GraphQL agent: an expert-calibrated LLM judge scores production conversations, a self-healing pipeline repairs failures, and the repaired trajectories are trained into a smaller model (SFT, then GRPO against the judge) that replaces the frontier model (source).
13.8x
lower cost per task, with higher accuracy than the best frontier model
A Qwen3-235B model fine-tuned on Tinker with investor-labelled data to do six daily information-filtering tasks (e.g. which news matters to a macro investor) (source).
1% → 99%
valid structured output after fine-tuning a small model
A proof of value on Red Hat OpenShift AI: GPT-4.1 generated synthetic training data, then small models (Phi-4 Mini 3.8B, Qwen 2.5 1.5B/7B) were fine-tuned for multilingual structured extraction from social media, deployable locally (source).