Enterprises constrained by data-residency rules often have to self-host large language models (LLMs). However, continuously adopting newer models without retiring older ones fragments the GPU pool and inflates serving costs. Researchers at Hugging Face tackled this by consolidating traffic from over 200 internal applications onto a single self-hosted model.
They first identified quality gaps through production error analysis along three axes: instruction following, function-calling, and internal task distribution. Instead of optimizing all objectives jointly—which can cause cross-domain reward interference—they trained a separate GRPO (Group Relative Policy Optimization) expert for each axis and merged them using two-stage SLERP (Spherical Linear Interpolation).
Each expert exposed a distinct failure mode: semantic collapse, over-calling, and verbosity hacking. These required domain-specific fixes. In non-reasoning mode, the resulting model outperformed a baseline about seven times larger in total parameters on the in-house Arena (69.6 vs. 65.8), instruction following (0.85 vs. 0.83), and function-calling (0.79 vs. 0.77), while also lifting general dialogue benchmarks.
The model now serves 50% of platform traffic—116 million requests per month—at a fraction of the serving cost. The authors note that the request mix shifts weekly, so the retraining loop must close gaps faster than apps drift. They also emphasize the importance of validating LLM judges, as judges tuned on the same traffic may measure mimicry rather than true service quality.