Affine (SN120)’s King Model Gets a 40.9% Throughput Boost From Cacheon (SN14)

Cacheon (SN14) has improved the inference throughput of Affine (SN120)’s King model by 40.9% after two weeks of optimization. Affine is also Cacheon’s first commercial customer, marking an important step toward bringing its technology to paying customers.

The optimization increased the model’s throughput from 375.4 to 528.9 tokens per second, outperforming the standard SGLang framework. This means the optimized system can generate more AI output in the same amount of time.

Affine (SN120) allows miners to improve AI models through reinforcement learning. Its focus includes reasoning, coding, and mathematics, with the best-performing model earning the title of “King.” The model optimized by Cacheon is based on Qwen3.6-35B, which uses only a portion of its parameters for each token.

Cacheon (SN14), meanwhile, focuses on making AI models faster and more efficient to run. Its network rewards participants who develop better inference systems without compromising the accuracy or quality of model outputs.

The improvement matters because running AI models requires considerable computing resources. Higher throughput allows operators to generate more output using available hardware, potentially reducing costs and serving more users. However, it does not necessarily mean that individual responses will be 40.9% faster.

The collaboration also highlights how subnets can support one another. While Affine works on developing more capable AI models, Cacheon focuses on making those models easier and cheaper to operate.

For Cacheon, securing Affine as its first customer is an early commercial milestone. The team plans to extend its optimization work to more models and customers, creating opportunities to generate revenue from its technology.

➛ Follow Cacheon (SN14) and Affine (SN120).