New playbook: when your own AI model pays. Read
Harvey (legal software, ~500+ employees) reported this on 20 Aug 2026 in company self-report (source). It belongs to the learning loop layer.
Harvey Tenet: Kimi K3 post-trained with asynchronous RL on long-horizon legal tasks, with rewards that also penalise wasted tokens; Harvey's stated aim is to let law firms train and own specialised models.
Harvey Tenet (Kimi K3 post-trained with Fireworks) completes almost twice as many held-out LAB tasks as base Kimi K3 (+9 pp all-pass rate). Separately post-trained Harvey models: a GLM-5.2 Review Table model (with Applied Compute) improves answer quality by 3.6 points and citation quality by 12.1 points versus the strongest baselines at roughly one-tenth the cost per cell; a Qwen3.8-27B firm-knowledge model (with Engram) cuts cost per query by 90%.
The number covers harvey's LAB hold-out and review-table tasks; research preview.
“Creating systems to allow law firms to build their own specialized models and own their intelligence,” Harvey 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).