DSpine
by Tencent · Model · Language & code models
promising
Speeds up Qwen3.8-27B by roughly 4-5x and edges out DFlash 2.
On one H100 with Qwen3.8-27B, DSpine hit ~150 tok/s vs ~129 for DFlash 2, about 16% faster
AccessOpen weightsNeeds~78 GBLicenseApache 2.0
Why we say so
Fahd Mirza, testing on one H100, measured about 150 tok/s against about 129 for DFlash 2, roughly 16% faster, and an earlier test went from about 33 to 156 tok/s. Tencent's own model card claims more accepted tokens per round than DFlash. All of this comes from one tester plus the maker, and it needs just over 78 GB of VRAM.
Evidence
- Nov 14Served on SGLang drafting 16 tokens per block; used just over 78 GB VRAM on an H100 80GB