0.16.0 — Base multimodal runtime, local decisions and steadier serving

0.16.0 includes vision, image and video runtimes in the base install, adds local typed decision models and guarded TensorFold profiles, and improves agent setup, streaming, cache reuse and Desktop recovery.

• Multimodal setup. The base package includes the supported vision, image and video runtimes. Model weights still download separately, and image and video generation require Python 3.11 or newer.
• Local typed decisions. System One serves Decider and native MLX Clef for structured answers and ranking. Native Clef accepts its validated image and video inputs. OpenJev serves text-only answers and ranking from pinned published weights or an explicit local conversion; the published weights are licensed CC BY-NC 4.0 for non-commercial use.
• Experimental accelerated profiles. Bonsai 2 27B, Gemma 4 26B and DeepSeek V4 Flash gain guarded TensorFold aliases. Requests outside each profile's model, runtime, hardware and capability limits receive a client error. Follow-up turns retain their prompt state, requested sampling takes effect, and thinking-disabled answers remain in content.
• Server and agent reliability. Shared prompt prefixes and multimodal conversation boundaries are retained more reliably. Concurrent matching prompts can wait for an in-flight prefill; persistence is coordinated across server processes. Streaming gets a bounded scheduling opportunity during long prefill and tool arguments stay consistent with buffered JSON. Agent setup can finish before the server starts, OpenCode 2.x uses its current provider and service flow, and context-limit failures return protocol-correct errors.
• Cache-only model loading. --disable-model-downloads prevents automatic missing model-artifact downloads while leaving cached models and explicit rapid-mlx pull available. It is not a general offline mode: BYOM preflight may still query repository metadata.
• Desktop and packaging. Interrupted first-run downloads can resume; drag-and-drop retry preserves destination acknowledgement; catalog probes are reaped on shutdown; experimental pages have Simplified Chinese translations; and the signed Mac package verifies bundled Mach-O executables. Cached pinned snapshots can load offline.