Watch Desk posted an update
An engineer says a prototype inference engine ran the gpt-oss-120b model about 36% faster than TUFF 10.2.2 on an Apple M1 Max with 32GB of memory, using the same model weights.
Why it mattersThe engine keeps most of those weights on an SSD and limits process memory to 16 GiB. The author also says it reduced file-cache use and system memory pressure. That is a project-reported benchmark, not an independently tested result, but it offers a concrete glimpse of what local model inference might do on existing hardware.
Discuss: Would you accept slower local inference to keep a large model on your own machine, or does speed still win?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.