Vela-1.0-Encoder-307M-CoreAI — Apple Core AI

A Core AI (.aimodel) build of vllm-sr/Vela-1.0-Encoder-307M for on-device inference on Apple Silicon.

Architecture: ModernBertForMaskedLM. Upstream ships PyTorch and no Core AI build.

Verified equivalence

The exported graph and the repository's own PyTorch loader were run on identical inputs and the raw outputs diffed:

measure	value
worst absolute difference	4.470e-07
output magnitude	1.00
relative	4.47e-07
minimum cosine	1.0000000000

Arithmetic equivalence only — no quality benchmark here: no F1, no accuracy score, and no claim against upstream's published numbers. The probes are the converter's default set; they exercise the graph rather than demonstrating the model doing its job.

Built by

ahmedandaloes — Apple Silicon builds (Core AI, MLX, GGUF) for models that ship none, each machine-checked against its own reference. ahmedandaloes.com