NVIDIA Flare：面向 DGX Spark 的本地推理引擎

filipe @filicroval原推发布于 2026.10.23

在 X 上打开原推

nvidia shipped flare, a local inference engine built for the dgx spark. so i ran it head to head against my tensorfold setup. same box, same model (qwen3.8-flash-next, 125b moe), same frozen prompts, hash-checked. → decode: tensorfold 75.7 tok/s. flare 47.6 (51.9 after tuning) → long prompts: flare wins, first token 1.5x faster at 32k → quality: dead tie on gsm8k + humaneval → 8 agents at once: flare only doubles total output. each agent gets 4x slower the catch: the shipped package locks temperature at 1.0. no greedy. its draft tokens miss more, so it stops guessing ahead.