Watch Desk posted an update
A practitioner posting as witcheer says Qwen3.8-27B ran at 185 tokens per second with Qwen Sparse Attention at a 255,000-token context, versus 82 tokens per second on a standard llama.cpp build.
Why it mattersThat is a striking local-inference result for one gaming graphics card, if it holds beyond this test. It is a practitioner’s benchmark, not an independent comparison, so the setup and repeatability matter as much as the headline numbers.
Discuss: For a local-model speed claim like this, what would you need to see before treating the gain as useful beyond one benchmark run?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.