wally models list --local is the live list. --cloud lists the cloud models on your account, and --all shows both.
This table groups models by publisher; the id you pull depends on the engine you want:
wally models pull <id>: the llama.cpp / GGUF build (macOS, Windows x64, Linux).wally models pull mlx-<id>: the MLX build (Apple Silicon only).MLX builds exist on Apple Silicon only; on any other platform they are hidden
from the catalog. The short aliases (qwen3, llama3.2, smollm2, …) still
resolve.
There are no Apple Neural Engine (ane-<id>) rows in this release. The NeuRT
engine that runs them is a private overlay, not in the public kit, and the
published Core ML repos hold directory trees rather than a downloadable
archive. The rows are commented out in src/catalog/catalog.rs under
TEMP(ane-cut) and come back together with real artifacts.
| Org | Families | Pull |
|---|---|---|
| Alibaba Qwen | Qwen3, Qwen3.8 | qwen3-0.6b, mlx-qwen3-0.6b |
| Meta | Llama 3.2 | llama-3.2-3b, mlx-llama3.2 |
| Gemma 4 | gemma-4-e2b, mlx-gemma-4-e2b |
|
| Hugging Face | SmolLM2 | smollm2-135m |
| Liquid AI | LFM2.5 | lfm2.5-230m, lfm2.5-350m, lfm2.5-1.2b, lfm2.5-2.6b |
| IBM | Granite 4.1, Granite 4.2 | granite-4.1-3b, granite-4.2-8b |
| NVIDIA | Nemotron | mlx-nemotron-nano |
| PrismML | Bonsai, Ternary-Bonsai | bonsai-1.7b, mlx-bonsai-1.7b |
| DeepGrove | Maple Preview | maple-preview, mlx-maple-preview |
Decision models return label probabilities rather than generated text. Use
wally decisions (alias wally decide), not wally run or a coding harness.
It supports yes/no (--ask), choices (--choice) and ordered levels
(--score); wally decisions --help shows the request-file and JSON forms.
--image FILE (repeatable, up to 8) adds PNG, JPEG or WebP images on the
hosted path only, for a model that takes them. On macOS a HEIC photo is
converted to JPEG first; elsewhere it is refused.
eve is the hosted decision model and the default, so -m is optional. Two
local checkpoints run the same questions on-device:
| Id | Backend | Pull |
|---|---|---|
clef-flash-9b |
llama.cpp (Q4_K_M) | wally models pull clef-flash-9b |
clef-flash-mlx |
MLX (Apple) | wally models pull clef-flash-mlx |
wally decisions --local -m <id> forces the local path and never touches the
network; a bare -m <local id> selects it automatically. On Apple, MLX
scoring needs the Swift runtime release that carries the decision callback;
the GGUF row runs on every kit today.