0.14.3 — Repeat image questions stop re-reading the image, and Desktop can finish small jobs

Asking a second question about the same picture no longer re-encodes it or re-reads the conversation. Personal Intelligence moves from answering to doing a bounded piece of work on your Mac — with the exact path, content and command shown before anything happens. The Community Benchmark says where it is while it runs. And the downloaded DMG opens correctly again on current macOS.

• A follow-up question about the same image skips the work it already did. On a qualified Qwen3.6 35B media workload the second turn reached its first token 54.8% sooner and finished 22.4% faster, and the singleton lane generated 35.4% more tokens per second. The media-aware prefix cache keeps the exact prior-turn media boundary so the image is not encoded again and the unchanged conversation is not prefilled again; its one-time storage cost measured 10–25 ms on short eligible first turns and nothing on long prompts. A separate fast path stops merging and repacking a cache whose shape is already right — across a 21-case suite the response hashes and peak memory were identical. Both are default-on, fail closed outside their qualified boundary, and keep an explicit rollback switch. Multimodal sampling also reaches parity with text: seed, top_k and min_p are no longer ignored by the media scheduler.
• Personal Intelligence can finish a small job, inside a fence you can see. A qualified local model can search and read local text, write a file you reviewed, move one file to recoverable Trash, and compile or run a bounded development command — enough to find a note in a folder, save a draft, or write and run a small program. It is not a shell. Desktop shows the exact path, content, command and arguments first; search and read can be approved for the session, while writes, execution and Trash need a fresh approval every time. Paths stay under your home directory, and hidden files, symlinks, credential stores, package descendants and oversized or non-text inputs are refused. Commands run with no shell in a macOS sandbox bounded on time, output, filesystem, network and processes, with compiler plugins and indirect argument files unavailable. Work lands in ~/Rapid Workspace when you do not name a destination.
• The Community Benchmark tells you where it is, and what it measured stays what it measured. Desktop shows live stages, pass counts, an ETA and the latest measurement while a run is going, then offers a direct route to run another model or repeat the same one. Contributor identity survives an app restart. My Results and Community views distinguish an exact total from bounded recent evidence rather than inventing a zero or a comparison. A completed result keeps its model, protocol, execution, Mac and contributor scope even if the on-screen selection changes afterwards, and packaged builds carry provenance that is stamped and verified before anything can be published — a modified source build still measures locally but fails closed at publication. DeepSeek V4.1 can now run the registered benchmark through its qualified serial runtime: on a 256 GB M3 Ultra the pinned 2-bit path measured about 31.09 tok/s short-case and 29.42 tok/s long-case.
• The downloaded DMG opens again. The 0.14.0–0.14.2 disk images stored two Finder view records in a nested binary-plist shape that could close the install window on current macOS, and crash and relaunch Finder on Sonoma. The layout is now generated deterministically using the flat dictionary records Finder writes itself, the release verifier rejects the old encoding, and the branded background and Applications drag target render correctly.
• A 32K prompt on GLM-5.3 Flash no longer has to fit in one forward pass. Native MTP prefills in bounded 1,024-token chunks instead of materializing the whole prompt at once; two consecutive real 32K HTTP requests completed on a 256 GB M3 Ultra without Metal running out of memory. This does not claim the checkpoint's advertised 1M context is usable. Alongside it: stopping rapid-mlx serve no longer hits the macOS 15 Python-finalization race that could hang or crash the process at shutdown, a cancelled DNS-pinned Desktop request cannot resume the same continuation twice, and a sandboxed command returns at its hard deadline even when macOS is slow to tear down a denied child.
• One more text alias, and the package has its real name. bonsai2-27b-2bit brings the catalog to 193 text aliases, 247 total; Bonsai 2 Hadamard MLX packs now load through a Rapid-owned loader for their packed Hadamard projections and inverse embeddings, qualified across text, streaming and image HTTP requests. The Python distribution also moves to the canonical rapid_mlx package name, keeping a compatibility shim so existing import vllm_mlx code goes on working.