Published one model at a time.
A model family appears here when it runs from a published build, not when it is on a plan. Each cell says which architecture, and since when. The table is drawn from the same catalogue the engine ships with, so it cannot get ahead of the software.
Architecture × model family
Loading the coverage table…
What a published build will actually do
Qwen2.5
EXPERIMENTAL0.5B, 1.5B and 7B base checkpoints on Apple Silicon. 32,768-token context, greedy decoding, quantised on the device while the checkpoint loads.
veizik run 7b "…"
Qwen3
EXPERIMENTALThe 1.7B base checkpoint. The adapter adds QK-Norm and drops the qkv bias; its logits were checked against a reference implementation to 4.3e-5 max absolute delta.
veizik run 1.7b "…"
Two lanes, and what is moving in each
Order, not dates. A lane opens when a package passes the gate, and the gate does not bend for a calendar.
NVIDIA CUDA
PREPARING- In flight — image and video families run on this lane internally: FLUX.1, Stable Diffusion 3.5, LTX-Video, Wan, CogVideoX.
- Gate — published when the package clears the same checks the Metal build cleared.
- Queued — language families on CUDA, after the image and video lane opens.
Apple Metal
OPEN- Landed — Qwen2.5 (0.5B / 1.5B / 7B), Qwen3 (1.7B), 32k context.
- In flight — the calibrated quality profile through convert; sampling beyond greedy.
- Queued — Llama and Mistral adapters; larger Qwen sizes.
Internally running is not the same as published, and this site never uses the first to imply the second. A family shown as in preparation has no download behind it today.