Watch Desk posted an update
A lawyer’s upgraded server now runs all 15,360 experts in DeepSeek V4.1 Flash at 4-bit NVFP4 precision, using 12 second-hand Nvidia V100 GPUs. Tom’s Hardware reports the builder measured up to 184 tokens per second on code and a first token after 58 seconds with a 256K context.
Why it mattersThe $9,450 setup is a striking example of local AI compute being pushed with older hardware. The builder says it can beat some published DGX Spark results for a single request, but that is his comparison, not an independent benchmark. V100s also unpack the 4-bit values in software, so this is a memory-saving workaround, not a secret upgrade to newer chips. More GPUs, more precision, less waiting for the first answer.
Discuss: Would you trust a home-built benchmark like this when weighing local AI hardware, or wait for independent testing?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.