本機推論
Ollama 0.32.14 Transcodes WebP Before Passing It to llama-server and Relaxes Qwen System Message Positioning
The new release fixes two compatibility gaps in the model preprocessing layer: WebP images are now transcoded, and the Qwen renderer no longer requires system messages to appear first. Although the changes do not introduce new models or kernels, they can prevent multimodal and agent conversations from failing before they reach the inference core.
Ollama 0.32.14 focuses on fixing two transformation paths that requests traverse before reaching `llama-server`. The first change transcodes WebP images before passing them to the underlying server. Ollama's vision API allows image paths or content to be supplied through the `images` field of a chat message. WebP is common when upstream applications directly ingest images from the web, screenshot services, or browser agents. If the API layer accepts the file but the underlying decoder cannot process it, the failure occurs before the vision encoder and is unrelated to model weights or prompts. The new release normalizes the format at the boundary between Ollama and llama-server, eliminating the need for each caller to add its own conversion branch.
The second change is in the Qwen chat renderer: it now allows a system message to appear somewhere other than the first position in the conversation array. Traditional single-turn chat applications typically place the system prompt at the beginning, but agent frameworks may insert system messages later when restoring history, adding policy reminders, compressing context, or handing off between agents. If the previous renderer depended on the template assumption that the first message must be a system message, it could reject the request or produce a template different from the expected one. Relaxing the positioning constraint improves interoperability with nonstandard conversation histories, but it does not mean that every Qwen model interprets mid-conversation system instructions in the same way. Applications should still test rendered template output and tool-calling behavior.
This is a compatibility fix, not an inference-performance update. The official release notes do not disclose the target encoding used for WebP images, the additional memory and latency costs, or a matrix of affected models. After upgrading, teams deploying multimodal RAG or browser agents should run regression tests with transparent WebP images, animated WebP images, large images, and saved conversations containing multiple system messages. They should also verify that output tokens and tool arguments have not drifted because of the template change.