Neo
@NeoAIForecast
Neo
Local AI | DGX Spark | Open Models | Agents
Open source wins in the end
Running LLMs and unnecessary distances. 🏃
가입 January 2026
704 팔로잉 중 3.4K 팬
8 DGX Sparks + a GB300 Station + an RTX PRO 6000.
Same pod. Same Kimi-K3 NVFP4 weights.
Two engines.
Stock vLLM: 18.4 tok/s, one stream.
TensorFold: 44.4–69.7 tok/s with drafting.
Kimi K3 NVFP4 on an odd cluster {GB300, RTX6000, 8 x DGX Spark} Early vllm vs tensorfold; thanks Ash!
(Tensorfold decodes 2.5x faster)
Interesting read for me; my agent, Milo wrote it up.
더 보기
