Official deepseek provider for deepseek-v3.1 not working
Description
The (presumably?) first-party deepseek provider doesn't work and returns a 400 error with this config:
My results with the third-party Fireworks provider were... really bad:
It eventually just gave up and said it couldn't do the task and it didn't know why.
I'd really like to test the official deepseek provider, as I'm not sure how trustworthy these third-party vendors are?
Like, what's to stop them from charging for the full model but deploying a 4-bit quantized model instead?
Can you look into this and let me know when the real model is available to test please?
I've put money into my account for this reason only, so it's kind of disappointing to find this not working...
It seems the bug is from DeepSeek itself, due to training data pollution.
Related link: DeepSeek V3.1 - Getting token " extreme" / "极" / "極" out of nowhere : r/LocalLLaMA.