0.16.1 — Safer local serving and more reliable model setup

0.16.1 tightens browser and Host access by default, supports dedicated embedding servers, and improves vision requests, model recovery, installation and coding-agent setup.

• Safer local serving. Foreign browser origins and unexpected Host headers are rejected by default. Explicit origin and trusted-host lists enable browser, LAN and reverse-proxy access. Service qualification works with the protected Host setting.
• Embeddings and vision. A server can run only an embedding model, with no chat model loaded. Anthropic image input is preserved, Gemma 4 12B image reading works again, and supported dense hybrid vision and text requests can batch.
• Model loading and recovery. Impossible loads are rejected before healthy models are evicted. Low-memory switches return retryable responses, while startup, capability and residency failures give a recovery command. Vision cache diagnostics identify the active backend.
• Agent and CLI setup. Codex setup accepts server-root URLs. OpenCode, Hermes, Pi, Qwen Code and OpenHands guidance and setup match their current clients. Model discovery suggests a command to serve a model. Installation through curl | bash preserves input for the installer.
• Serving reliability. Recurrent hybrid KV quantization remains selective, shared prompt state is counted once, and MCP startup and teardown are more robust. Required tool choice works with native XML templates, Anthropic custom tools require an input schema, and speech uploads keep their container suffix.
• Desktop and sharing. Desktop restores vision capability when resuming a model, allows a custom model-download endpoint, and keeps update offers clear of model controls. Pool sharing shows status and first-run download progress, prevents prompt-cache persistence, and fails closed on TLS errors.