ModelsBenchmarks ReleasesDocumentation
Releases  ·  Engine

Apple Metal

The published lane. One runtime, four models, and an installation that verifies itself before it runs.

LANE OPEN FIRST PUBLIC RELEASE 2026.09.20
Current release

2026.09.20 — first published package

Version
2026.09.20 · catalogue 2026.09.16
Operating system
macOS on Apple Silicon (arm64). Nothing is compiled on your machine.
Devices
M1, M2, M3, M4 and M1 Ultra declared. Exercised on M4 Pro, M1 Ultra, M1 Max and M2 Pro.
Models
Qwen2.5 0.5B / 1.5B / 7B (base) and Qwen3-1.7B (base). Other architectures are refused, not attempted.
Context
32,768 tokens per core, prompt and generation combined.
Artifact
10 MB, 14 files, checksummed end to end.
Build path
No interpreter anywhere in the build or install path.

What changed in the engine

Compatibility

Fixes carried into this build

Integrity

CHECKSUMSEvery file in the archive verified on install14 / 14 OK
INSTALLATIONThe runtime verifies itself and its components before it will runfail closed
CATALOGUESigned, and doctor reports the resultsigned

Known limits, stated plainly

veizik.limits · 2026.09.20
  • Quantisation is uncalibrated. Output is coherent but below the calibrated profile; on the 0.5B model it is noticeably so.
  • Greedy decoding only. No temperature, top-p or top-k yet.
  • You bring the checkpoint. The runtime reads your Hugging Face cache; it does not fetch weights for you in this release.
  • Base checkpoints only for the registered models. Instruction-tuned variants are not registered.
  • Four models in this package. Other architectures are refused rather than attempted.
History

Engine entries on this lane

Loading…