Apfel vs TurboOCR

Apfel is a Swift CLI that does something Apple didn't: it exposes the on-device LLM baked into every Apple Intelligence-enabled Mac as a proper OpenAI-compatible local server running at localhost:11434. Any app that speaks to Ollama's API — LM Studio, Continue, OpenWebUI, your own scripts — can now route requests to Apple's FoundationModels framework without modification. The feature set is more complete than most indie wrappers: streaming responses, tool calling with MCP support, file attachments, an interactive chat mode, and a debug SwiftUI GUI for inspecting token flow. Inference is fully on-device with no API keys, no telemetry, and no cost beyond electricity. On an M-series Mac, it runs at native Apple Neural Engine speeds — typically 40-80 tokens/second depending on the model variant active. The catch is real: you need macOS 26 Tahoe (currently in beta) and Apple Intelligence enabled. But for the tens of millions of Apple Silicon Mac users who already qualify or will soon, this is the quiet unlock of a model they already own. The "your Mac already has a free LLM" framing is resonating — the repo hit 3,500 stars in days.

TurboOCR is a C++20 OCR server that uses CUDA and TensorRT to process documents at speeds that make Python-based OCR look like a fax machine. The headline number: 270 images per second on FUNSD form datasets with approximately 11ms single-request latency — roughly 50x faster than PaddleOCR's standard Python implementation. It uses PP-OCRv5 models (the same underlying tech as PaddleOCR) but squeezes them through TensorRT FP16 optimization for GPU inference. The server exposes both HTTP and gRPC interfaces from a single binary and handles PDFs natively with four extraction strategies: pure OCR, native text layer extraction, hybrid verification mode, and a "best of both" fallback chain. PP-DocLayoutV3 handles layout detection across 25 document region classes — useful for structured documents where you need to know that a bounding box is a table cell vs. a header vs. a figure caption. A Prometheus metrics endpoint tracks throughput, latency, and GPU memory in real time. Deployment is Docker-first: TensorRT engine compilation happens automatically on first startup. The catch is it requires Linux with an NVIDIA Turing GPU (RTX 20-series minimum) and driver 595+, so it's not a laptop tool. But for enterprise document automation — invoices, forms, medical records — the throughput-to-cost ratio is hard to beat.