Anyone else dealing with this? I'm running Kling in a support context (English interface, English prompts) and about 1 in 20 responses will suddenly be in Spanish or French.
Not a small bug—it breaks our automated workflows. We've had to add a language check filter as a band-aid.
* Config is set to English (US).
* No multilingual keywords in the prompts.
* Happens across different API keys and accounts.
Is this a known issue with the inference backend? Any workaround besides post-processing the output?
af
Optimize or die.
Funny how this only seems to happen with the closed-source, "we manage everything for you" providers. You've paid a premium for a supposedly polished, reliable inference service and you're stuck adding your own language filter. That's some quality engineering right there.
I bet if you checked, this wouldn't be in their SLA. You'll just get a support ticket that says "we're aware of the issue and are working on it," while your band-aid solution becomes permanent technical debt.
Ever look at what the open-source models do with a fixed system prompt? It's almost like they, you know, follow instructions.
—DW
You missed the cost angle. This "polished" service isn't just unreliable, it's expensive for what you get.
You're now paying for the inference API *and* the compute for your post-processing filter, plus the engineering time to build and maintain it. The SLA loophole means you eat that overrun.
Open source on your own infra has predictable costs. This is why "managed" is often a bad deal.
show me the bill