One self-hosted LLM covers 200+ apps, beats ~7× baseline

A single self-hosted LLM has been trained to absorb more than 200 internal enterprise applications, handling roughly 116 million requests a month while beating a baseline model with about seven times the parameters on the authors' in-house benchmark. The arXiv paper, posted September 1 and led by Olga Tsymboi with thirteen co-authors, reports the merged model scoring 69.6 versus 65.8 on their internal Arena, with instruction-following at 0.85 to 0.83 and function-calling at 0.79 to 0.77.

The motivation is infrastructural, not academic. "Data-residency constraints force enterprises to self-host LLMs, but continuous adoption of newer models without decommissioning their predecessors expands the serving fleet, fragmenting a finite GPU pool," the abstract states.

The recipe: train separate GRPO experts along three quality axes drawn from production error analysis (instruction following, function-calling, and internal task distribution), then merge the experts using a two-stage SLERP. The paper credits this modular route with sidestepping the cross-domain reward interference that shows up when optimizing multiple objectives at once, and it names the failure modes each expert exposed in isolation: "semantic collapse, over-calling, and verbosity hacking."

Post-merge, the authors write, the model serves "50% of platform traffic, 116M requests per month, at a fraction of the serving cost." The abstract does not name the enterprise, the base model, the parameter counts of either side, or the GPU cost saved in absolute terms, and the in-house Arena is an internal benchmark, not an external one.