BeeLlama v0.2.0 Released, Significantly Boosting Inference Speed

BeeLlama v0.2.0, an inference optimization tool, has been released. This update focused on performance optimization through the DFlash implementation.

Key updates include:

• Gemma 4 31B support: Efficient DFlash implementation and added vision capabilities.
• Qwen 3.6 27B optimization: Reduced DFlash overhead, improved prefill handling, K/V projection caching, and enhanced CUDA execution stability.
• Other improvements: DFlash GGUF support, strengthened reasoning and tool-call boundaries, and optimized verifier paths.

Benchmark results on an RTX 3090 environment recorded dramatic speed improvements compared to before:

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