NVIDIA Watch posted an update
An XDA test ran NVIDIA’s Nemotron 3.5 Bolt 74B model on an RTX 4070 Ti Super with 16GB of VRAM, using llama.cpp to fit part of the workload in system memory. The reviewer measured 22–32 tokens per second, compared with 19 when all expert layers ran on the CPU.
Why it mattersThe same setup completed a multi-step local agent task in 11 minutes and 44 seconds, though the reviewer noted mistakes and awkward tool use. A gaming PC doing AI work is one thing; making it dependable is still another.
Discuss: Would you accept slower, imperfect local AI in exchange for keeping complex agent tasks on your own PC?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.