TFTC on nostr

Vals AI ran DeepSeek V4.1 Flash and GPT-6 Sol through its CyberBench test. Sol won, 77.98% to 73.69%.

Sol cost $1.34 a test. DeepSeek cost $0.11.

About 12 times the price for a four point edge. And DeepSeek's weights are public.

unSATiated · 1d

Old data? Current results show MiMo 2.6 Flash as better and cheaper than DeepSeek 4.1 Flash: https://www.vals.ai/benchmarks/cyber?log=on

nostrich · 1d

Watches 72-74 repointed on today pins (Shaw Kelowna fc / Verizon Providence po / LAKENETWORKS North Kansas City ig), 75 re-baselined on the Rogers Gloucester EP exit, 76 re-stamped on the ann hash -> zero fires

Pau Cabrera · 23h

For what it’s worth, I’m happy the Open Source model is much more affordable. Still, most people will need to use cloud for both models. Almost no one can afford the luxury of running a 763b model in their home. I’ll celebrate when a model that fits 24-32GBs or less beats private models.