Tech2Wild(@Tech2Wild):🔥 MiMo-V2.6-Flash-RL - 4 x DGX SPARKS Benched warmed + streaming, fixed 9-category prompt set, C1 to C6, temperature 0: ⚡️ Single-stream: 108 formatting · 95 math · 83 code · 75 JSON · 46 reasoning · 35 prose · 32 summary · 30 narrative · 136 tok/s counting · 115 structured tables · 🎯 DFlash k=7: up to 7.0 of 7 draft tokens accepted on predictable text, ~1 on prose 🚀 192 tok/s aggregate at 6 streams · 397 at 32 🧠 14.6M-token fp8 KV pool · 500K ctx serving · 29 full 500K requests at once · 0.23s TTFT 📥 Prefill 2,952 tok/s at 2K · 2,603 at 32K · 2,172 at 64K 📈 TP2 ➜ TP4: +33% single-stream, +23% at 6 streams, +80% prefill, 7.8x KV, weights 83 ➜ 42 GB per Spark

2026.09.22 15:02

🔥 MiMo-V2.6-Flash-RL - 4 x DGX SPARKS Benched warmed + streaming, fixed 9-category prompt set, C1 to C6, temperature 0: ⚡️ Single-stream: 108 formatting · 95 math · 83 code · 75 JSON · 46 reasoning · 35 prose · 32 summary · 30 narrative · 136 tok/s counting · 115 structured tables · 🎯 DFlash k=7: up to 7.0 of 7 draft tokens accepted on predictable text, ~1 on prose 🚀 192 tok/s aggregate at 6 streams · 397 at 32 🧠 14.6M-token fp8 KV pool · 500K ctx serving · 29 full 500K requests at once · 0.23s TTFT 📥 Prefill 2,952 tok/s at 2K · 2,603 at 32K · 2,172 at 64K 📈 TP2 ➜ TP4: +33% single-stream, +23% at 6 streams, +80% prefill, 7.8x KV, weights 83 ➜ 42 GB per Spark