The latest release candidate for Ollama, v0.32.14-rc0, includes two focused improvements for self-hosted users. First, llama-server now automatically transcodes WebP images before sending them to the model. This means you can use WebP-format images in your prompts without manually converting them, even if the underlying model does not natively support that format.
Second, the Qwen renderer has been updated to tolerate system messages that are not placed at the very beginning of a conversation. Previously, some Qwen models would reject or mishandle chats where a system prompt appeared later in the message history. With this change, self-hosted users can use more flexible conversation structures without running into errors.
