9 November: Featherless models wait while they load instead of failing
Chats with less common Featherless models now wait while the model loads, tool calls from more model families work, and the list of usable models matches what Featherless serves today.
People who chat with less common Featherless models see fewer failed first messages and a short loading notice instead. The picker lists only models Featherless serves today. No action is needed.
Loading models. Featherless loads less common models on demand. The first message to such a model used to fail with "temporarily at capacity", and the chat switched to a fallback model at once. ZeroLog now shows that the model is loading, waits up to 75 seconds and asks again. If the model is still not ready, the fallback works as before.
Tool calls. Gemma and Mistral models write tool calls as plain text. These calls are now recognised and run. Tool-call markup and reasoning that some models put into the answer no longer appear in the visible text. Models with strict message templates get a message order they accept, and models with a very small context window answer without tools instead of failing.
Current list. Featherless has cut the models it serves from about 22,000 to about 10,000. The usable models in the picker now match that, at about 10,100. Models Featherless no longer serves can no longer be selected, so a chat does not fail on its first message.
No action is needed. If a chat with a Featherless model failed today, send the message again.