github gwenn-ha-dev/Siliconed v0.30
Siliconed 0.30

one hour ago

The first edition named after a beach: Rhossili, three miles of sand at the end of the Gower peninsula, in Wales. Every version now gets its own, somewhere on the Atlantic coast between Ireland and Galicia; the name shows at the right of the window's bar and in “About”. From 0.21 to 0.30, the jump says the break: a new engine, and a new window.

Core AI

Siliconed used to carry its own Metal kernels, tuned chip by chip and measured on each Mac before its first render. From 0.30 it computes on Core AI, the runtime Apple ships in macOS 27: Apple writes and tunes the kernels for every chip, the GPU and the Neural Engine alike, and Siliconed keeps what is its own — the models, their memory, their proof. macOS 27 is required; on macOS 26, 0.21 remains the last version.

  • Apple's kernels, on every chip. Each model is a few Core AI functions written in Swift and the loop that calls them: no kernel of ours to maintain, no table of chips, nothing measured before the first render. A new Mac, a new macOS, and Core AI brings what suits them (The engine).
  • The Neural Engine works on every model. Its share of each block is found during the render, step after step, and moves to whichever unit finishes first; the next render starts where the last left off. The Standard can use it too, its bf16 weights rounded once to f16; the engine times both ways once on this Mac and keeps the faster. Precision › Exact keeps the GPU alone, in the model's own arithmetic.
  • The published files, as they are: the publisher's bf16 for the Standard, published int8 for the Compact, dequantized inside Core AI's functions. An installation is a download, with no preparation after it, and a model's two versions live side by side.
  • Under the memory limit, from the first render. Every stage is planned to the byte before anything is allocated; a model larger than the limit keeps the weights that fit and reads the others from disk while it computes; what Core AI holds for its functions is read in flight, every millisecond; a large image is decoded in bands; what cannot fit is refused up front. The limit (¾ of the Mac's memory minus 2 GB by default) is set from 4 GB to the memory minus 4 GB, in Settings › Macs (Memory).
  • LoRAs computed beside the product they adapt, never merged into the weights, on both versions: merged, a LoRA's small change was lost in the int8 codes and partly in bf16.
  • Compiled once. Core AI compiles a function for its shapes: the first render at a new size, in a version not yet run or with a LoRA of a new rank compiles first, and the Chain says so (“Preparing for this Mac, once…”). Every render after it starts at once. Settings › Macs weighs what was compiled and erases it (“Compiled for This Mac”); an update of the app erases it by itself.
  • Checked in the arithmetic of the weights: the Standard against diffusers and transformers in bf16, the Compact against ComfyUI's int8 products, each session's memory against its plan (VERIFY.md).
  • The bits belong to the system: the same seed gives the same picture to the eye, no longer necessarily the same file byte for byte from one macOS to the next.
  • Times: Numbers. On the Mac mini M5 Pro, the Standard is as fast as in 0.21 or faster; the Compact is slower (1.1 to 1.5 ×): 0.21's int8 kernels were our own, and on a GPU that computes bf16 as fast as f16 the Compact now buys memory, not time.

The app

  • A new window, in three planes. The image is the window's background, from edge to edge. On the left, one column, the chain from the idea to the button: the Text and Images boxes (the Prompt Lab), the encoders, the DiT with its settings, the VAE, and Generate pinned at its foot; it folds (⌥⌘R, or the toolbar's button) and its edge drags. On the right, time: the queue as ghost tiles above the render under way, then the history, which widens into a grid over the image (⌥⌘G) (The window).
  • Glass: the two columns are glass over the image. Fitted, the image holds whole in the room they leave; zoomed in, it slides under them to the window's edges, and so do the curtain and a series' grid. Settings › General › Window › Glass makes them opaque.
  • The chain detaches into a floating window of its own (the toolbar's second button); closing it docks the chain back.
  • The Chain panel draws the render's memory in time: one curve over the whole render, the memory limit across, a mark where each stage begins, its weights kept on hover; a render on another Mac comes with its curve too. The image's details (⌘I) are a popover of the ⓘ under the image, each stage with its time, memory and unit; the Chain inspector (⌥⌘I) gives one line per stage, with the GPU's and the Neural Engine's shares of its arithmetic.
  • Import a model or a LoRA: from the menu, the Models window's “Import…” button, or by dropping a .safetensors on the window. The file is recognized by its tensors, copied as it is and read in its author's format; a non-distilled base model is refused, and the app says why. From a Mac computing on another, the file is sent to that one (silicontrol import).
  • LoRAs show their trigger words when the file gives them: under the LoRA, a click puts them in the prompt.
  • The Prompt Lab has nothing left to download: the writer is Qwen-Image-2.1's encoder as published. When the Standard does not fit the memory limit, the app offers the Compact to write on this Mac.
  • Detail on the five models: Anima, Krea 2 and MOSS-Image now offer “Detail: Normal · More · Most”, each at a setting found for it.
  • MOSS-Image reads long prompts, up to 1,024 tokens like the others, and starts from an image too.
  • Precision replaces the Accelerators switch (Settings › General): Exact keeps the GPU alone, in the model's own arithmetic; ANE, the default, lets the Neural Engine take the work it does faster on this Mac.
  • The history grid selects with a lasso, and its head exports, copies or removes the selection. Every series of the session is kept: its cells are ordinary images of the history, under a series tile that folds. The sheet is an export (Export Sheet…, Copy Sheet), and Seeds is a series too, its cells still rendered in one pass.
  • The zoom is bounded in the image's pixels: up to 16 points per pixel, drawn as squares past one point a pixel.
  • Space moves the components aside (Image › Move the Components Aside, ⌥⌘F): the columns and the bar under the image go; Space or Esc brings them back.
  • A new version is said at launch: the app asks GitHub for the latest number, and a newer one shows in a small panel, with a button that downloads its disk image. Settings › General › Updates turns it off.
  • Downloads carry no token: every repository the app reads is open. A gated one linked from Hugging Face offers no file; its sheet says to download it from its page and import it.
  • The window is at least 1140 × 680.
  • silicontrol: set glass [on|off]; set updates [on|off]; set memory; compiled [erase]; sheet g2 <file|folder>; grid --wait --out D writes the sheet there (--no-sheet skips it); rack fold|unfold|top|dit|detach|dock; remove g2 and reuse g2; uninstall legacy; set preview [on|off]. siliconed uninstall acts only with --yes.

What goes

  • The Light versions (4 bits) and GGUF reading: the “Open in Siliconed” link from a Hugging Face page now offers .safetensors files only.
  • Faster Rendering (Beta): every evaluation runs at full size, the recipe checked against the reference.
  • “Measure This Mac…”, the capability map and the diagnostic report: Core AI chooses for the chip, and what the engine needs it reads during the render.
  • The Prompt Lab's own window, and its fold (⌥⌘L): one column from the idea to the button now. silicontrol lab open|close|detach|attach go with it.
  • Full screen, replaced by Space; Statistics and the details bubble, replaced by the Chain inspector and the image's details; Add Sheet to History, replaced by the export.
  • The old library is not carried over: models are reinstalled from the app, and the Models window offers to free the room the 0.21 files held (Free…).

Don't miss a new Siliconed release

NewReleases is sending notifications on new releases.