DSpine boosts Qwen3.8-27B to 94-132 tok/s in llama.cpp, 1.75x over DFlash2
pbaylies · x · 2026-11-09
Tencent Hunyuan released DSpine, a speculative decoding method for Qwen3.8-27B, and developer ItsmeAjayKV has ported it to llama.cpp with pre-built GGUF quantizations (upstream llama.cpp doesn't support it yet, so you need to clone the fork).
Benchmarks:
- Baseline: 28.2 tok/s
- DFlash2: 50.6 tok/s
- DSpine: 94–132 tok/s
DSpine leads DFlash2 by 1.75x, driven by predicting more tokens per verify step on math and code (7.3–7.6 vs 5.7–5.8). Full details are on the model card and GitHub repo.
More from Infra
- Nvidia is the second-largest position in SPY, yet the largest position inside a SPY — dr1337 · 2026-11-09
- OpenAI CFO says datacenter expansion won't slow, stressing local community ties — i_dg23 · 2026-11-09
- 92T tokens over 1.54M DAU: the agent-era math feud over token burn — teortaxesTex · 2026-11-09
- Gary Marcus backs Ed Zitron: US data center capacity additions halved in Q4 2025 — GaryMarcus · 2026-11-09
- Analysts: Nvidia could stabilize near $155 if Huang is cleared, but an exit over security violations would hurt far more — GaryMarcus · 2026-11-09
- Texas gubernatorial candidate vows moratorium on all data center approvals — Traditional-Chip8339 · 2026-11-09