Qwen3.6-35B-A3B
RTX 3090 Boosts Qwen3 MoE Prefill by 35%, Cutting Local Inference Costs
A custom cudaMemcpy tuner lifts Qwen3.6-35B-A3B prefill on one RTX 3090 from 345 to 466 tokens/s.
1 article tagged with this topic
A custom cudaMemcpy tuner lifts Qwen3.6-35B-A3B prefill on one RTX 3090 from 345 to 466 tokens/s.