More DS4.1 TP4 improvements on 4x DGX Spark 🚀 - Prose decode: 85 tok/s at c1 (was 69) - Boot: ~3 min - Prefill: ~4.8k tok/s at 32k–128k, 4.3k at 262k - KV cache: ~6.5M tokens, 1M context - Same weights, same quality (qeval unchanged)

Tokens/sec lied to me. Same 30 tasks, 4x DGX Spark: DeepSeek-V4.1-Flash: 61 tok/s, thinks 39k tokens → 671 s GLM-5.3-Flash NVFP4 (effort high): 37 tok/s, thinks 5k tokens → 324 s Same answers. GLM done 2.1x sooner. At effort max GLM thinks like DeepSeek and loses. Agentic coding (6 repo tasks via OpenCode): DeepSeek 172 s, GLM 243 s. Little thinking there, raw speed wins. Rule: GLM on high for chat, DeepSeek for tool loops. Full numbers, launcher, harness:

Heads up for DGX Spark owners: capping clocks with `nvidia-smi -lgc 0,2200` keeps the GB10 cooler and the clocks stable but it does NOT persist across reboots or driver resets. Wrap it in a systemd timer (OnCalendar=*:0/10) or your "capped" box quietly drifts back toward 3 GHz.

I've published this DeepSeek 4.1 Flash recipe for 4x DGX Spark (mainly clone of Mia's one with tweaks and experimental libs etc). Prose C1: 45-> 55 tps. C16: 134 -> 277 tps.