v0.5.0

Models you don't own, if you want them: a Cloud tab for bring-your-own-key providers — GMI Cloud (MiniMax M3 and M2.7, plus speech and music), OpenRouter, and NVIDIA's free developer tier. Off until you add a key, and never picked for you: your agents still fall back to a local model unless you choose otherwise. OpenMontage, the open-source agentic video studio, gets a Set up button in Settings — one click installs it to ~/OpenMontage and links this app in as a $0 provider for images, video and 3D, so its productions render on your Mac and your swarm. Qwen Code's daemon now starts with a bearer token, so adding a workspace works. Silicon Node gets its own download and section on the site.

v0.4.0

Pi joins the engine roster — earendil's coding agent, running on your models with the app's whole toolbox bridged in. Settings grows Your other AIs: one-click Connect buttons that wire Claude Desktop, Claude Code and Codex to your local models, so your subscription stops burning tokens on the heavy work. Swarm invites now hand each person their own revocable key instead of the shared one.

v0.3.0

The swarm grows up. A new Swarm tab is a command center for every machine — memory, GPU, what's serving, what's queued — with start/stop/switch controls, context windows up to 256K, per-model visibility switches, and a Test button that proves the whole path in one click. Every request any engine routes through the app is now logged with timings and token counts, inspectable down to where the time went. Friends on your Tailscale network can join your swarm with a six-digit code, like pairing headphones. Under the hood the model gateway learned to speak both thinking-control dialects, diagnose empty answers instead of shrugging, advertise itself in a discovery file, and queue behind a busy GPU when you ask it to. Plus: quantizations are named in every picker, and Codex and Qwen Code got the full-window button.

v0.2.1

Everything the app needs now ships in the box: the llama.cpp engine (with expert streaming) and the Node.js runtime for the harness chat are bundled, checksummed, and documented in THIRD_PARTY_LICENSES.md — a fresh Mac works with zero terminal commands. Plus: missing 3D model weights download with one click, every generated image gets a Make-it-3D button, the 3D tab can draft its source image from a text description, and all status messages are written in plain language.

v0.2.0

The 3D toolkit: turn an image into a 3D model with TRELLIS.2 (textured, local), Hunyuan3D 2 (clean geometry in seconds, native MLX) or a remote LATO.2 service — with an interactive viewer, Fusion 360 handoff, and full Claude/ChatGPT control via MCP. Qwen3.8 27B joins the curated catalog with figures read from the file itself. Generated filenames now use 8-character suffixes so large batches cannot collide.

v0.1.0 — first public build

A native macOS menu bar app that works out what a local model will cost in
unified memory before you download it, then proves the prediction.

Install

1. Download Silicon Optimizer.dmg, open it, drag the app to Applications.

2. This build is signed locally rather than notarized, so macOS will block it on
   first launch. Clear the quarantine flag once:

   xattr -d com.apple.quarantine "/Applications/Silicon Optimizer.app"

   Or build from source, which avoids the prompt entirely.

3. Install a runtime — the app finds it automatically:

   brew install llama.cpp

What it does

Predicts weights, expert pool, KV cache and compute buffers from the model's
real architecture rather than estimating from file size, then tells you whether
your Mac can run it and how fast. If it will not fit, you get a ranked list of
what to change with the exact saving of each.

A bundled MCP server lets Claude or ChatGPT drive the model you already have
loaded, instead of starting a second copy.

Measured on an M3 Max, llama.cpp b10280

	Predicted	Measured
Qwen3-Coder 30B-A3B, Q4_K_M, 16K	20.30 GB	20.24 GB
Generation	89 tok/s	89 tok/s
Expert streaming, 24 of 128 experts	5.20 GB	5.32 GB

Expert streaming runs the same 30B model in 5.3 GB instead of 19 GB, with
identical output — only residency changes.

Known limits

• Calibrated on one machine. The built-in benchmark corrects for yours.
• Not notarized, hence the quarantine step above.
• Expert streaming needs a llama.cpp built with the expert-paging patch; the
  app detects the flags and hides the option when they are absent.
• MLX runs but has no catalog entries yet — point it at a model directory.
• Split-file models resolve correctly but the multi-part download is untested.

Full detail: https://optimize.zamasu.dev