ShapeLearn发布Qwen 3.8 27B完整量化版本，性能达BF16基准99.63%

ShapeLearn团队发布了Qwen 3.8 27B模型的完整量化版本，相比此前的ShapeLearn-Lite版本提升了性能表现1。该完整版本包含五个不同量化级别的模型，在6种GPU配置下的测试中均实现了质量与速度的帕累托最优1。

其中GPU-5量化版本达到BF16聚合基准的99.63%性能，吞吐量在59.2至93.7 token/秒之间波动（具体数值取决于GPU型号），在RTX Pro 6000上达到90.4 token/秒，在RTX 5090上达到93.7 token/秒1。GPU-4版本在内存占用11.0 GB的情况下实现了98.72%的BF16性能1。

研究引入了MTP和DFlash2两种推测解码方法来进一步提升效率1。其中DFlash2的吞吐量提升倍数为1.34至2.10倍，MTP为1.28至1.66倍1。KLD评估基于约500万个token的数据集，涵盖长上下文和Agent任务场景1。该研究论文已被EMNLP Industry Track接收1。

The ShapeLearn team has released a full quantization suite for the Qwen 3.8 27B model, building upon an earlier lightweight version to deliver improved performance across multiple optimization levels 1. The complete offering includes five distinct quantization variants designed to achieve optimal quality-speed tradeoffs across six different GPU configurations 1. The GPU-5 quantized version achieves 99.63% of the BF16 baseline performance while delivering throughput between 59.2 and 93.7 tokens per second depending on the GPU type, with peak speeds of 93.7 tok/s on RTX 5090 and 90.4 tok/s on RTX Pro 6000 1. The GPU-4 variant maintains 98.72% of BF16 performance at 11.0 GB memory consumption 1.

Beyond quantization alone, ShapeLearn integrated two speculative decoding methods—MTP and DFlash2—to further amplify throughput 1. DFlash2 achieved throughput improvements of 1.34 to 2.10 times the baseline, while MTP delivered 1.28 to 1.66 times baseline throughput 1. The evaluation methodology employed a dataset spanning approximately five million tokens, incorporating both extended context windows and agent-based tasks 1. The research has been accepted to the EMNLP Industry Track 1.