Veizik for Apple Silicon — 2026.10.02-r5
Veizik for Apple Silicon. One signed package that reads supported safetensors checkpoints directly and runs eight language models on the machine you already have. 16.1 MB installed.
This page records the state as of 2026-10-04 and is not updated as the catalogue moves. For what runs now, see models.
What this build does
- `veizik run --ignore-eos`: generate exactly the requested number of tokens, like llama.cpp --ignore-eos, so fixed-length comparisons are like for like.
- `--tokens N` now prints exactly N tokens and says so on the summary line.
- Faster long-context decoding on more Macs: Qwen3-1.7B takes the long-context path from 8K tokens on M2 Pro, M1 Max and M1 Ultra, and Qwen2.5-1.5B gains a faster path on M2 Pro.
- `veizik serve` adds a TLS gateway, per-tenant quotas, a multi-node router, API keys with scopes and single sign-on, a usage ledger, request ids, tool calls and a service command.
- `veizik serve` now keeps each model loaded in a resident engine: one load per server, not per request, with model and feature licences checked when it loads and on every request.
- Licences can now cover individual models and features.
- 16.1 MB installed — 16,130,971 bytes in 125 files, summed from the published disk image itself, not from a build tree.
- Signed with a Developer ID and notarised by Apple; the image carries SHA256SUMS over its other files.
Models in this release
| Model | Task | Context | Status |
|---|---|---|---|
| Qwen2.5-0.5B | Language | 32,768 | Available |
| Qwen2.5-1.5B | Language | 32,768 | Available |
| Qwen2.5-7B | Language | 32,768 | Available |
| Qwen3-1.7B | Language | 32,768 | Available |
| Qwen2.5-0.5B-Instruct | Language | 32,768 | Available |
| Qwen2.5-1.5B-Instruct | Language | 32,768 | Available |
| Qwen2.5-7B-Instruct | Language | 32,768 | Available |
| Qwen3-1.7B-Base | Language | 32,768 | Available |
- Platform
- macOS on Apple Silicon (arm64)
- Install size
- 16.1 MB, 125 files — no dependency to install first
- Verification
- The file name, size and SHA-256 are printed on this page and on the release page, and the two agree. The disk image is signed with a Developer ID and notarised, and carries SHA256SUMS covering its other 49 files, so `shasum -a 256 -c SHA256SUMS` checks the contents after you open it.
- File
- veizik-metal-2026.10.02-r5.dmg
- Size
- 5,981,740 bytes
- SHA-256
- 3bfaa0fcb3f812cb7c03275e26f4ac2d713e604c1fdcdddcace8f7c9cf4d8b4e
Verify, install, check
$ shasum -a 256 -c SHA256SUMS $ ./install.sh $ veizik doctor
Full steps are in the documentation.
Fixed issues
- `veizik serve` no longer grows in memory when a client polls /ready or /metrics.
- `veizik serve` no longer holds on to memory when it refuses a streaming request that also asks for tools.
- Sign-in tokens signed with ES256 are verified about 15 times faster, and a recently verified token is not re-verified on every request.
Known issues in 2026.10.02-r5
- `--tokens 1` is refused: the first token comes from reading the prompt, so the smallest output is two tokens.
- Compared with 2026.10.02-r3, `veizik run` wall time is one decode step shorter for the same --tokens value, because --tokens now counts the first token.
- If the engine cannot start on your Mac it says so and exits rather than running anyway. This only happens on configurations that restrict what an application may do with memory, which a stock macOS does not; if you see it, support@veizik.com.
- Headless enrolment is not enabled: every device is activated through a browser.
- The benchmark board carries no figures for this release yet. All four engines changed, so earlier figures are not reused; the board returns when this package has been measured.
- `veizik serve` handles prompts up to 4096 tokens by default; longer requests are refused.
- On long prompts `veizik serve` can be slower than `veizik run`: the resident engine does not yet use the faster long-context path.