you just leaked your api key.
#Empty response from model
halcyon summit
already deleted. do not post keys or things that look like keys
tame hollow
so i need to repost?
halcyon summit
but again, you put a text to text model as the vision model
nope
I have the data
tame hollow
but again, you put a text to text model as the vision model
gemma-3n-e4b-it is vision model, wdym with text to text?
halcyon summit
oh, that variant is, yes. there are so many gemma e4b models. the listing i was referencing had it as text to text from openrouter non instruction tune variant.
Your current debug share is for profile gacha, and the visible failure is still going through the main custom model path:
provider: custom
model: deepseek-v4-pro-precision
The important line in your screenshot is:
attaching 1 image(s) natively (model supports vision)
That means Hermes is not using the auxiliary NVIDIA vision route for that image turn. It is attaching the image directly to the current main model request, and that main endpoint is returning empty content.
If you want Hermes to force the auxiliary vision model instead of native image attach, add this to the active profile config:
~/.hermes/profiles/gacha/config.yaml
agent:
image_input_mode: text
auxiliary:
vision:
provider: nvidia
model: google/gemma-3n-e4b-it
timeout: 500
Do not keep the API key in config.yaml; keep it in .env as NVIDIA_API_KEY.
Then restart Hermes and test again.
If it still fails after that, send a fresh hermes debug share from immediately after the failed attempt. The thing we need to see is whether the log now shows the vision request resolving to provider=nvidia / model=google/gemma-3n-e4b-it. Right now the visible failure still looks like native image routing into the main custom model, not a confirmed NVIDIA auxiliary vision failure.