Supported models

areno currently supports the following checkpoint families:

The table describes CUDA model adapters. Apple Silicon support is determined by the installed mlx-lm or mlx-vlm version and the checkpoint’s model type; consult Apple Silicon and MLX before assuming that the same family is available on both backends.

Family	Notes
Llama-style dense decoder models	Dense causal decoder checkpoints with Llama-compatible layouts.
Qwen3 dense	Qwen3 text checkpoints.
Qwen3-MoE	Routed expert checkpoints with Areno MoE kernels.
Qwen3.5	Dense text and Qwen3.5-VL image checkpoints.
Qwen3.5-MoE	Qwen3.5 routed expert text and VL checkpoints, including qwen3_5_moe layouts, with Areno MoE kernels.
Bailing MoE Linear v2	Local model adapter for Bailing MoE Linear v2 checkpoints.
Ling / Bailing MoE V3	bailing_hybrid checkpoints with softmax/KDA attention, sparse MoE, and activation checkpointing.
Gemma4	Gemma4 text and conditional-generation checkpoints. Native Gemma4 and Gemma4 Unified processors support image, audio, and video inputs for serving and training.
MiniCPM-family adapters	MiniCPM-family text and vision adapters used by the local training stack.

Important

Model support means the checkpoint can be loaded through an Areno model adapter. Some model families may support inference before every training or save path is fully optimized.

On MLX, AReno delegates model construction and weight sanitization to mlx-lm/mlx-vlm and saves a native MLX checkpoint. Backend support must be validated independently, especially for quantized and multimodal checkpoints.

For the common media message format and runtime behavior, see Multimodal inputs.