A developer running Qwen3.8-27B on an RTX 3090 GPU increased token generation speed from 82 to 382 tokens per second through software optimizations, achieving a 4.6x improvement without hardware upgrades. The gains came from speculative decoding techniques that process multiple tokens simultaneously rather than sequentially. The custom inference engine used three methods: multi-token prediction tuning, a drafting system called DFlash2 that proposes seven tokens at once, and a lookup mechanism that identifies repeating text patterns to fill prediction slots. These approaches reduced the bottleneck of repeatedly fetching model weights from memory, which typically limits performance more than raw processing power. Accuracy on math benchmarks remained at 96.5 percent during testing. Speed improvements occurred in stages, with intermediate results showing 114 tokens per second after MTP tuning, 138 after adding the lookup drafter, and 381 during large document processing. The optimizations required storing context in a more compact format, which doubled initial document loading time but enabled a 138,696-token context window. While everyday chat saw speeds around 133 tokens per second, the techniques are most effective for repetitive text generation tasks. The developer noted trade-offs including reduced available context slots and slower initial document ingestion, but emphasized these methods could extend the useful life of existing hardware for specific workloads.