Shapelearn Qwen 3.8 27B (13.1 GB VRAM) (byteshape.com)

Shapelearn has launched its full ShapeLearn Qwen 3.8 27B models, significantly enhancing performance over the previously released ShapeLearn-Lite variants. The new models showcase considerable improvements in quality-speed metrics, with all five models topping the performance charts in multiple GPU comparisons. GPU-5 comes highly recommended for its impressive performance, achieving 99.63% of the BF16 aggregate benchmark score, whereas GPU-4 offers a smaller footprint with only a slight trade-off in throughput. Notably, the release integrates advanced features like Speculative Decoding with MTP and DFlash2, which enhance throughput across all tested models. DFlash2 generally delivers faster results but requires more memory and doesn't support image inputs, making MTP the better choice for memory-constrained environments or when multi-modal support is necessary. The comprehensive benchmarking against competitors highlights that ShapeLearn models consistently lead in both accuracy and speed, marking a significant step forward for developers and researchers in the AI/ML community focused on optimizing model efficiency and performance.