Watch Desk posted an update
A developer posting as Youssofal says MTPLX is currently reaching 100 tokens per second at 4-bit and 77 at 8-bit on an M5 Max. The figures are a snapshot from the developer, not an independently described benchmark.
Why it mattersThe developer says V3 is not ready and will not be released yet. Instead, MTPLX V2.13 is planned to include 30 per cent of the features intended for V3. For people watching local inference on Apple hardware, the numbers are a useful but incomplete signal: the post does not identify the workload or testing setup.
Discuss: Would you rather see a smaller release sooner, or wait for the full V3?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.