T-Tech's Merged Qwen3-32B Handles 116M Requests, Beats Qwen3-235B

T-Tech is now running more than 200 internal applications on a single Qwen3-32B, the group reports, covering 116 million requests a month at what the abstract calls 'a fraction of the serving cost': 2.8 to 3.9x lower per token on input and output, and 4 to 9x lower for the largest previous services.

The base is Qwen3-32B with a Cyrillic-adapted tokenizer. The baseline for comparison is Qwen3-235B-A22B-Instruct-2507, roughly seven times larger by total parameters. On the in-house arena stratified to production traffic, the merged model scores 69.6 against the larger baseline's 65.8; on instruction following, 0.85 to 0.83; on function-calling, 0.79 to 0.77.

Rather than train against every objective at once, which the paper says introduces 'cross-domain reward interference,' the authors train a separate GRPO expert per axis, then merge them with two-stage SLERP weight interpolation. Each expert's reward, they write, 'exposes a distinct failure mode, namely semantic collapse, over-calling, and verbosity hacking, each requiring a domain-specific fix.'

Semantic collapse is minimal but formally correct answers. Over-calling is tool invocations when the model is uncertain. Verbosity hacking is response inflation to game a length reward. Joint multi-objective training, the authors report, requires 1.7 times the budget to match the merged model.

An open-weight checkpoint, T-pro-it-2.1, is on Hugging Face, trained without internal data but with the same recipe; the authors argue this validates that the recipe, not proprietary data, drives the gains. The paper joins a run of self-hosted deployment work in our open-source tracker this week.