Efficient LLM Inference for Diverse Muslim Languages and Cultures

We propose a competition on efficient LLM inference for languages widely used across Muslim communities. Participants may use any efficiency method or inference engine, including fully custom systems, while passing multilingual and culturally grounded quality gates. The H100 divisions use Gemma 4 31B; the NVIDIA GB10 (DGX Spark) division uses Gemma 4 24B MoE. We evaluate Arabic, Persian, Urdu, Bengali, and Indonesian on 1$\times$H100, 2$\times$H100, and GB10 under low-, medium-, and high-load Poisson traffic. A request is successful only if TTFT $<10$ s and decoding speed is $\geq20$ tokens/s on H100 or $\geq15$ tokens/s on GB10; systems are ranked by throughput contributed only by successful requests.