Benchmark compares speculative decoding engines for Qwen 3.6 27B on RTX 3090

A user benchmarked five llama.cpp forks and variants using the Qwen 3.6-27B model on a single RTX 3090 to evaluate speculative decoding performance.

• ik_llama with ubergarm-tuned MTP achieved the highest narrative speed at 63.9 TPS, while its ngram+MTP variant reached 87.8 TPS for code generation.
• beellama using DFlash with an independent draft model recorded the highest code throughput at 96.8 TPS but suffered from a high TTFT of 504ms.
• Mainline llama.cpp offered the lowest TTFT (288ms) and maintained consistent speed across context lengths, unlike forks that showed significant degradation.
• Spiritbuun provided strong consistency with only -9% context degradation, while LUCEBOX performed poorly across metrics.

The results highlight trade-offs between raw throughput, time-to-first-token, and context stability when selecting an engine for local inference.