@Alibaba_Qwen Qwen3.8-Flash-Next on two @NVIDIAAI DGX Sparks just got better 🚀 It inherited some of the improvements i did on the single-spark recipe and a little extra more! Default recipe 👇 ・Multi-turn fix: follow-up replies 5x faster (3.25s → 0.63s), after tool calls 2.55s → 1.98s ・Code up to +27% faster (333 → 422 tok/s at 8 streams) ・~56 tok/s prose, ~76 code (1 stream), up to ~226 prose / ~422 code (8 streams) ・Optional bit-exact decoding across both nodes ・Cached-token reporting in the API ・Safer memory default: no more running both nodes at under 1 GiB free New opt-in lane on vLLM 0.30 (./start-v030.sh) 🌟 ・Follow-ups in 0.28-0.40s, and repeated prompts hit the cache (0.35s vs 2.7s) ・Same decode speed, steadier: fixed a vLLM 0.30 default that made speed swing ±15% between runs ・One small patch instead of seven ・Optional FP8 KV: 2.6M tokens of context, 3/3 needles at 200k Tested for stability for a few hours! I will keep testing this over the next days and probably make the vLLM 0.30 script the default one in the next version of the recipe ✌️