1.0.0 — Agent chats and new models

Start a persistent agent chat with rapid-mlx chat --harness, with tools, approvals and image attachments. Desktop also offers harness chats through an experimental setting that is off by default.

• Work with an agent. Terminal chats show measured model-loading progress. Network clients require approval and terminal commands inherit a restricted environment; the network policy is heuristic and terminal commands are not sandboxed.
• Try new models. LFM3 1.7B text and LFM3-VL 2B image models join as pinned 8-bit aliases, alongside Ornith 1.5 122B-A10B in 4-bit. MiniMax-H3 generates video with synchronized audio from text or a first image; it needs 160 GB+ unified memory and BF16 weights, and defaults to the 8-step Turbo LoRA.
• Share compute in Desktop. The experimental setup now has three steps: add one QuickSilver provider key, choose or download a model, and start sharing. Credits uses that same key; the separate read-only key and Live Pool tab have been removed. Public shares forward inference, health and model discovery routes, while local control routes remain private.
• Keep chats on track. Desktop accepts up to eight images per message and can optionally release idle models. Improved parsing keeps reasoning and tool calls in their intended fields. Qwen3.8-27B TensorFold share nodes accept tools but still reject response formats, grammar and media requests.