[HYPER] all inputs have 0% cache hit

[sppfly]

Description

I'm using DeepSeek V4 pro/flash through Charm Hyper and seems like the cache hit rate of these two models are all 0%. Gemma 4 as well.

This happens both with Crush CLI and VSCode Copilot, so I guess there's some problem with the proxy.

Version

0.76

Environment

No response

[nds-stack]

hyper.charm.land is the fastest way to get poor, and it's absolutely ridiculous. What's the point of them mentioning cache write/read prices if tens of millions of tokens are burned with 0 cache hits?

[andreynering]

Hey @sppfly!

This was a temporary issue that we fixed today. DeepSeek V4 Flash & Pro will get cache now.

Thank you a lot for reporting and for your patience! Hyper is still in Beta, so this is really useful for us.

[sppfly]

@andreynering Hi thanks for replying. Unfortunately the issue still persists. the number of cached tokens is still the same no matter with crush or other agents.