Zazen 0.1.0-alpha.4 (experimental)

Experimental alpha. Zazen runs commands and edits files on your PC (it asks before risky steps), its AI makes
mistakes, and things change without notice. Use it on files you have backed up.

Windows 10 (22H2) or 11, 64-bit, with a dedicated graphics card with 8–16 GB of video memory and 16 GB of RAM.
NVIDIA is tested (RTX 3070 Laptop, 8 GB). AMD Radeon and Intel Arc are supported but not yet tested on real
cards: reports welcome.

Install

1. Download ZenHarness-Setup-0.1.0-alpha.4.exe (282 MB) and check it against the .sha256 file
   (Get-FileHash .\ZenHarness-Setup-0.1.0-alpha.4.exe).
2. It isn't code-signed yet: if Windows says "Windows protected your PC", choose More info → Run anyway.
3. Start Zazen and click Set up this PC. It finds your graphics card and downloads, once: the llama.cpp runtime
   for it (CUDA, HIP, Vulkan or SYCL builds of PrismML's fork), Ternary Bonsai 2 27B in the size that fits (5.9 or
   7.2 GB), its vision file, and offline Wikipedia (14 GB, optional). Every file is checksum-checked; downloads resume.

Your chats and settings from an earlier alpha are kept.

What's new since alpha.3

• Any GPU vendor, 8 to 16 GB: hardware detection picks the runtime, model file, context (64K / 96K / 128K),
  cache and where image reading runs. Every choice can be changed in Settings → Hardware.
• A much smaller installer (282 MB, was 820 MB): runtimes and models download on first run for your hardware.
• Agent harness: code changes need a passing test run before "done", and the program must actually run (a page
  loads, a server answers); loop guards that no longer deadlock; drafts kept when a write is refused; small tasks
  skip the project ceremony.
• Answers: charts, maps, tables, timelines, fact panels and diagrams inside the answer.
• Documents: six themes across Word, PowerPoint, Excel and PDF, with charts and tables.
• Reliability: a crash in the window shows the error with Reload instead of a blank screen, and is logged to
  ui_errors.log; the window no longer flashes white while loading.
• Updates: installed copies check GitHub Releases and offer the new installer.
• Renamed to Zazen and relicensed under Apache 2.0.

Known limits

• 3D games and visual design work poorly on this model; small apps, scripts, data work and fixes in existing
  projects work best.
• Expect 15–30 tokens a second on an 8 GB card, slower as long tasks fill the context.

Testing

zen-harness-tester-kit-0.1.0-alpha.4.zip + TESTER_GUIDE.md: unzip the kit, open Claude Code in it and say
"Read TESTER_GUIDE.md and run the test plan in it."

Zen Harness 0.1.0-alpha.3 (superseded by 0.1.0-alpha.4)

Alpha for testers — use this instead of alpha.2. Windows 10/11 x64, NVIDIA GPU with 8 GB+ recommended.

Install

1. Download ZenHarness-Setup-0.1.0-alpha.3.exe and check it against ZenHarness-Setup-0.1.0-alpha.3.exe.sha256
   (Get-FileHash .\ZenHarness-Setup-0.1.0-alpha.3.exe).
2. It isn't code-signed yet: if SmartScreen says "Windows protected your PC", choose More info → Run anyway.
3. On the Models page choose which models to download (≈22 GB for all three, each SHA-256 checked):
   Ornith 1.5 9B · Qwen3.6 35B-A3B REAP-48 (mixture of experts) · Ternary Bonsai 2 27B (1-bit; setup also fetches
   PrismML's llama.cpp build it needs). The app starts on the first one you downloaded.

Prefer to build it yourself? See docs/BUILD_FROM_SOURCE.md.

Testing

zen-harness-tester-kit-0.1.0-alpha.3.zip + TESTER_GUIDE.md: unzip the kit, open Claude Code in it and say
"Read TESTER_GUIDE.md and run the test plan in it."

Changes since alpha.2

Found by watching a 1-bit model build an app live:

• read_file never hides part of a file. A file under 600 lines but over the 8,000-character result limit was
  "shown whole" and then cut — the model was told it had the whole file and could never see its end. Files are now
  paged honestly, and every read starts with the file's size and exactly which lines are shown
  ([server.py: 331 lines, 12,426 characters — showing lines 137-279; … continue with offset=280]).
• Old write calls in the history no longer carry text a model can copy into new files (the cause of a wiped file).
• Context masking keeps earlier pages of a paged file; guides and the tool catalog are no longer cut off.
• Fewer false alarms: paging a file, testing bad inputs with different bodies, and re-checking a server after a POST
  are no longer treated as loops or repeats; the PRD check finds the project's PRD from tests/.
• Bigger GPUs are used automatically: on a 12–16 GB card the built-in profiles get a full-precision cache, more
  context (up to 128K) or all MoE experts on the GPU — the 8 GB behaviour is unchanged.

License

PolyForm Shield 1.0.0 — free to use and change,
personal or business, but not to provide a competing product. Source-available, not OSI open source.

Zen Harness 0.1.0-alpha.2 (superseded)

Superseded by v0.1.0-alpha.3 — it fixes file reading for long files and more. Please use that one.

Alpha for testers. Windows 10/11 x64, NVIDIA GPU with 8 GB+ recommended.

Install

1. Download ZenHarness-Setup-0.1.0-alpha.2.exe and check it against ZenHarness-Setup-0.1.0-alpha.2.exe.sha256
   (Get-FileHash .\ZenHarness-Setup-0.1.0-alpha.2.exe).
2. It isn't code-signed yet: if SmartScreen says "Windows protected your PC", choose More info → Run anyway.
3. On the Models page choose which models to download (≈22 GB for all three, each SHA-256 checked):
   Ornith 1.5 9B · Qwen3.6 35B-A3B REAP-48 (mixture of experts) · Ternary Bonsai 2 27B (1-bit; setup also fetches
   PrismML's llama.cpp build it needs). The app starts on the first one you downloaded.

Testing

zen-harness-tester-kit-0.1.0-alpha.2.zip + TESTER_GUIDE.md: unzip the kit, open Claude Code in it and say
"Read TESTER_GUIDE.md and run the test plan in it." It drives the app through its local API, runs the expense-tracker
build benchmark on each model and the standard benchmark battery (IFEval, BFCL, MMLU-Pro, GPQA), and writes a report.

Changes since alpha.1

• Models are downloaded during setup instead of on first run; built-in tuned profiles for them.
• Harness: refuses writes that would copy the history placeholder into a file; never runs a tool call cut off at the
  output limit; a write that would wipe most of a file can't be forced by resending; more thinking while debugging
  (failing tests, tracebacks); asks for comments that explain the why.
• Per-model llama.cpp builds (llama_dir), strict chat templates, mixed K/V cache types.
• Personas and the Telegram bot are no longer part of the product.

License

PolyForm Shield 1.0.0 — free to use and change,
personal or business, but not to provide a competing product. Source-available, not OSI open source.