Commentary on X
as you get into the 100-200k context, it of course slows down a little bit, but still gettin 313 tokens/s on average is pretty damn great
5:54 AM · Jul 15, 2026@casper_hansen_ yeah the 'some tasks' qualifier is doing a lot of work. chained flows that are mostly tool calls and waiting don't compress at 541 t/s, you're just burning the $54/hr clock faster.
7:33 AM · Jul 15, 2026@casper_hansen_ its more like 100 tps for real tasks and a few sessions in parallel. Guess it is marketing
7:52 AM · Jul 15, 2026@casper_hansen_ 43 new contributors this cycle means sglang updates are coming in hot
7:00 AM · Jul 15, 2026Combined views
4.8K
3 posts, first seen 7h ago
23 likes2 comments