The VIDRAFT team documented its winning entry to the Fast Gemma Challenge, reaching 510.58 tokens per second at a perplexity of 2.39 on a single NVIDIA A10G GPU. The approach combined sliding-window attention limited to 188 tokens, speculative decoding with a 7-token drafter, and synthetic warmup procedures to stabilize measured performance. The team restricted itself to “quality-neutral” optimizations that would not compromise model quality, and built on shared community contributions such as pre-quantized weights.
