QVAC vs llama.cpp

Key differences

QVAC's inference engine, Fabric, is a fork of llama.cpp and tracks it upstream, so the two share a lineage and much of their model support. llama.cpp is distributed as source and binaries to compile and call directly. QVAC wraps Fabric behind a typed API in JavaScript, TypeScript and Python.

llama.cpp ships text generation, speech synthesis, embeddings, multimodal input, an RPC server that splits one model across machines, and Android and SwiftUI example projects. Model downloading and caching, per-device memory ceilings, a packaged mobile SDK and task-level APIs fall outside its scope, and an application built on it usually supplies them.

The llama.cpp tree does not cover OCR, neural machine translation, image generation, or LoRA fine-tuning on mobile GPUs, and QVAC provides all four, translation through the Bergamot engine from Mozilla. llama.cpp does support training; what Fabric adds on top is LoRA specifically, masked-loss instruction tuning, and mobile GPU support.

This page compares llama.cpp v0.3.0, released 25 August 2026, against QVAC 0.18.2, meaning the SDK together with the Fabric inference engine at v10297.1.1. Every row was checked against the project's own documentation and release notes on 4 September 2026. Both projects move quickly, so check the current release before you make a decision on either one.