Original title: From Production Traffic to Post-Training: Building a Self-Hosted LLM That Covers the Corporate Request Mix
Olga Tsymboi, Dmitrii Stoianov, Ramil Latypov, Danil Taranets +10 · September 2026
Companies that run their own artificial intelligence models often juggle too many systems at once. This wastes computer chips and drives up costs. To fix this, researchers analyzed real workplace errors and trained one unified model to handle jobs from over 200 internal applications. Training an AI on everything at once can cause goals to clash. Instead, the team trained separate expert models for three specific skills: following instructions, using external digital tools, and tackling company tasks. They then fused these models together using a two-stage blending technique. The finished model outperformed a baseline roughly 7 times larger by total parameters on internal tests. It now processes 50% of the platform's traffic, handling 116M requests per month at a fraction of the cost.
Imagine organizing a busy restaurant kitchen where the chef has to prepare many different dishes but only has one set of cooking stations. Each dish requires different skills—some need precise timing, others need specific ingredients handled carefully. If the chef tries to do everything at once, some dishes get ruined due to oversight. Here, the researchers break down the chef's skills, train specialists for each type of dish, and then combine these experts into one chef who can handle all dishes efficiently.
Companies that must keep data on their own servers waste GPU resources juggling multiple AI models. This research shows how to build one model that handles all their tasks, cutting costs and simplifying operations.
The abstract does not establish how generalizable this approach is beyond the specific enterprise context, as it relies on internal benchmarks and production traffic patterns.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
From Production Traffic to Post-Training: Building a Self-Hosted LLM That Covers the Corporate Request Mix — Olga Tsymboi, Dmitrii Stoianov, Ramil Latypov et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.