Kaffeine ARM64 preview 001: test 68k apps in QEMU

On this page

Kaffeine now has a package anyone can try. ARM64 preview 001, published on 2 November 2026, contains a prepared AROS system, Kaffeine’s 68k JIT, the Moira interpreter, a small original test program and the matching sources. You download it, extract it and start it inside QEMU: there is no installation on an existing AROS system.

Until now the Kaffeine results were mine alone: Personal Paint on AROS ARM64 and 68k applications running with the JIT. This preview has a different purpose: letting the project meet machines, applications and habits other than mine. It is a pilot for testers, not a compatibility release.

What the package contains

  • a clean AROS AArch64 image matched to the rest of the package, booting under QEMU emulating a Raspberry Pi 3 (1 GiB of memory, a single translation thread);
  • two icons under Kaffeine → KaffeinePreview, Test-JIT and Test-Moira, which run the same original 68000 program with the two backends;
  • the Start.command launcher with its Python front end: Python 3.9 or later and QEMU are installed separately;
  • matching public sources, exact catalog and CPU-core revisions, DTB inputs, build archives and original license notices. No proprietary application, ROM or NDK is included.

Every session starts from a fresh writable disk, so the supplied base image stays unchanged. Anyone who wants to try their own program can use the optional --import-app workflow described in the README (it requires mtools), importing the application into an isolated disk.

What was actually verified

Runtime qualification was done on macOS ARM64 with QEMU 11.0.0, with a 1024×768, 24-bit desktop. Both icons ran the test program, returned zero and saved a 105-byte file with the expected text, checked again after shutdown. The logs confirmed the two actual backends, and the JIT counters are nonzero: 29 blocks, 72 translated instructions, 247 entries.

The checks also covered startup from the extracted package, normal shutdown, a real one-minute timeout (the session ends at about 60 seconds with code 2, as specified), application import and local reports. The metadata-only report contains a single file; the opt-in one adds three short log tails. Nothing is uploaded automatically. In continuous integration, the six Python host matrix jobs and the Windows filename check passed, but the 37 tests use synthetic inputs and do not replace a full virtual machine boot.

The package is 451,581,722 bytes and the published SHA-256 checksum is b72876e4b8a64f889a3e9ed633cb88edfe12d9a0a9f754b114e49fcf1dd42dcd: verify it before extracting.

The limits, plainly

  • The included test is a small original program: it does not demonstrate compatibility with arbitrary applications.
  • Actual Linux and Windows boots, physical Raspberry Pi hardware, HVF acceleration and 3D are not yet qualified.
  • The full Lua gc.lua case does not finish within 600 seconds; no JIT defect is demonstrated, but it remains incomplete. CopyMem work is deferred.
  • The JIT is experimental. If an application fails, the failure alone does not prove a JIT defect, so it is worth repeating with Moira and noting which backend was used.
  • It is not an installer for an arbitrary AROS system. The long-term goal remains integration into mainline AROS.

One scope note: the recent fix for typed Workbench startup of 68k programs is in the repository but, according to the README, is not part of preview 001.

How to report a useful result

Repeatability matters. Include the application version, the exact steps, the launch method, the observed backend and the comparison with Moira. Generate the local report, inspect its contents and only then submit it through the community test form. Logs are optional.

The preview can be downloaded from the official release page together with its checksum file; the launch, import and reporting guide is in the preview README. This is how living retrocomputing works: first a reproducible result, then other eyes trying to break it.

Cover: conceptual illustration generated with OpenAI; it is not a screenshot of the software. The technical information comes from the release page and the attached QUALIFICATION.txt file, consulted on 4 November 2026.

Sources and documentation

AUTHOR

IT specialist, developer and systems engineer with a long history across code, Linux servers, retrocomputers and e-learning platforms. On AndroidLab he brings a technical, pragmatic eye: less brochure smoke, more attention to infrastructure, usability, privacy, updates and the real consequences of manufacturers' choices.

Leave a Comment