BeeLlama v0.2.0: Faster Inference with DFlash

BeeLlama v0.2.0 has been released. The core of this update is improved inference performance through a major overhaul of the DFlash implementation.

Key update details are as follows:

• Efficient DFlash implementation for Gemma 4 31B and Vision feature support.
• Qwen 3.6 27B performance optimization: reduced DFlash overhead, improved prefill processing, drafter K/V projection caching, and safe CUDA execution support.
• DFlash GGUF architecture support and various bug fixes.

Benchmark results on an RTX 3090 environment showed dramatic speed improvements over the previous version:

• Qwen 3.6 27B: up to 163.9 tps (a 4.40x improvement over the previous version).
• Gemma 4 31B: up to 177.8 tps (a 4.93x improvement over the previous version).