EXPERIMENTALMetal
0.5B, 1.5B and 7B published on Apple Metal
First published package. The binaries that were measured are the binaries in the archive.
Qwen2.5
Three base checkpoints run on Apple Silicon from the published cores. You supply the weights; the engine reads them from your own Hugging Face cache and prepares them on the GPU as they stream in.
EXPERIMENTAL
Apple Metal · since 2026-09-20
NVIDIA CUDA · planned
What each size costs on the device
Checkpoint size is what you already hold on disk. Device memory is what the engine allocates after quantising — measured on an Apple M4 Pro on 2026-09-20 with the published cores, not estimated.
| Size | Short name | Checkpoint | On device | Layers | Hidden | Heads (Q / KV) | Context |
|---|---|---|---|---|---|---|---|
| Qwen2.5-0.5B | 0.5b | 942 MiB | 327 MiB | 24 | 896 | 14 / 2 | 32,768 |
| Qwen2.5-1.5B | 1.5b | 2,944 MiB | 883 MiB | 28 | 1,536 | 12 / 2 | 32,768 |
| Qwen2.5-7B | 7b | 14,525 MiB | 3,923 MiB | 28 | 3,584 | 28 / 4 | 32,768 |
In all three cases the engine reported 0 MiB copied verbatim — every weight is rebuilt on the GPU, none is passed through.
The command, and what it prints
# the weights are yours to fetch $ huggingface-cli download Qwen/Qwen2.5-7B $ veizik run 7b "The capital of France is" device : Apple M4 Pro model : qwen2.5-7b H 3584 layers 28 checkpoint : 4 shard(s), 14525 MiB weights : 3923 MiB (0 MiB copied verbatim) attention : 28 heads, 4 kv heads, ctx 32768 Paris
Greedy decoding only. Temperature, top-p and top-k are not implemented in this release; the same prompt gives the same answer every time, which is also why the output is byte-identical across machines.
The context is a hard edge. Prompt and generation together may total 32,768 tokens. Over that the core prints a message and exits 1 — it does not silently drop the beginning of your prompt. VZ_CTX lowers the window if you want a smaller one.
Nothing is written. There is no converted artifact and no weight cache. Quantisation happens on the GPU during load, every time.
What has been checked, and on what
- Architecture support
- qwen2 — complete: import, preparation, tokenizer and chat template, plus numeric, generation and performance validation.
- Status
- EXPERIMENTAL. Every published size was proved by running it, not by registering it.
- Devices exercised
- 1.5B and 7B: M4 Pro, M1 Ultra, M2 Pro (7B), M1 Max (1.5B). 0.5B is the correctness-gate model.
- Targets declared
- M1, M2, M3, M4 and M1 Ultra.
- Determinism
- Generated text is byte-identical across three different Macs on the full prompt board (12/12).
- Context
- 32,768 tokens — the runtime's measured limit, not the architecture's. Verified 2026-09-16 by recall, and the release gate checks that the runtime and the catalogue agree.
- Quality
- Below the calibrated profile. The quantiser runs without calibration in this release, which is noticeable on the smallest model. veizik convert will produce the calibrated profile.
How this family got here
What became possible for this family, and when. Engine-wide changes live on the lane page.
registry 2026.09.16Context
Context opened to 32,768 tokens
Verified by recall, across every size.
packagingPack format
Component compatibility is checked, not assumed
A mismatched installation now fails visibly instead of running with the wrong parts.