DeepSeek V4.1 Flash 1017 local deployment math: 763B params need ~760GB, two DGX Sparks fall short
The author estimates ~760GB to load the 8-bit 763B model, more than two 128GB DGX Sparks, with 4-bit still ~380GB, and lists four benchmark wins over Opus 5 against seven losses.
TWO DGX SPARKS WON'T RUN THIS MODEL. I DID THE MATH DeepSeek V4.1 Flash 1017 hit Hugging Face today. MIT license, 763B parameters > about 761B of those weights are stored in 8-bit > so you need roughly 760 GB just to load it > one DGX Spark has 128 GB of unified memory > two linked Sparks give you 256 GB, and NVIDIA rates that pair for models up to ~405B > even a 4-bit quant lands around 380 GB Jensen hand-delivered a Spark to Elon at Starbase last year. it wouldn't load this either where it does beat Opus 5, per DeepSeek's own table: > DeepSWE v1.1: 77.0 vs 74.0 > Automation-Bench: 60.5 vs 50.3 > Agents' Last Exam: 35.4 vs 28.6 > Apex Bench: 72.1 vs 65.8 Opus 5 wins the other 7 rows, including HLE (56.3 vs 42.6) and Terminal-Bench 4.0 (51.8 vs 42.3) save this before you order hardware
This post quotes another post: x.com/0xspikez/status/2111494068773753112