Qwen model corrupt output stream

[John6666]

Maybe this issue?

Is the serverless API completely broken and unreliable?

I couldn’t find any real info anywhere so I subscribed with PRO to test whether I could use Llama 3.3 70b for my (small) app since the API seemed fine with mistral nemo. Unfortunately, I get garbage responses about half the time. Example: 07 every 08: this is low10 an08: this is not a good 07/1.0780780:00 to 1.irectional 07: this is 08:0780780: this is boot1: this is 0780:0000: this is 07:00:0780780780: this is 07: this is 07:00: this is 08: this is 08: this is 1: this is 07: this is 07:00: …