https://halfpennymac.com/blog Benchmarks and guides from Halfpenny Mac: real numbers from Apple silicon Mac minis running local LLMs, Swift builds, and always-on AI agents. en-gb https://halfpennymac.com/omlx-vs-llamacpp-qwen38-27b-mac-benchmark https://halfpennymac.com/omlx-vs-llamacpp-qwen38-27b-mac-benchmark Wed, 19 Aug 2026 09:00:00 GMT We benchmarked oMLX against llama.cpp running Qwen3.8-27B on the same 24GB M4 Pro Mac mini: 48% faster generation, a 4.6x faster cached prompt, and one surprise about MTP on Apple Silicon. https://halfpennymac.com/qwen38-27b-m4-mac-mini-benchmark https://halfpennymac.com/qwen38-27b-m4-mac-mini-benchmark Tue, 18 Aug 2026 09:00:00 GMT We ran Alibaba's new Qwen3.8-27B on an M4 Pro Mac mini with 24GB, four days after release. Real llama.cpp benchmarks, the GGUF quant to pick, the macOS wired-memory fix, and when you need a Mac Studio instead. https://halfpennymac.com/hermes-vs-openclaw https://halfpennymac.com/hermes-vs-openclaw Mon, 13 Jul 2026 09:00:00 GMT An honest comparison of Hermes (Nous Research) and OpenClaw, the two open-source always-on AI agents. Channels, models, memory, setup, and what it takes to run one 24/7. https://halfpennymac.com/swift-build-benchmark-m4-mac-mini https://halfpennymac.com/swift-build-benchmark-m4-mac-mini Mon, 08 Jun 2026 09:00:00 GMT We benchmarked Alamofire, Vapor, SwiftLint, and swift-algorithms on an M4 Mac mini. 19s for Alamofire, 64s for Vapor's 1,940-file build. Here are the numbers. https://halfpennymac.com/qwen35-gpt-oss-m4-mac-mini-benchmark https://halfpennymac.com/qwen35-gpt-oss-m4-mac-mini-benchmark Sun, 31 May 2026 09:00:00 GMT We benchmarked Qwen 3.5 9B, Qwen 3.5 4B, GPT-OSS 20B, and Qwen 3 4B Thinking on an M4 Mac mini. Reasoning models hit 30+ tok/s, OpenAI's first open model runs at 9 tok/s. Here are the numbers. https://halfpennymac.com/qwen3-llama3-m4-mac-mini-benchmark https://halfpennymac.com/qwen3-llama3-m4-mac-mini-benchmark Fri, 29 May 2026 09:00:00 GMT We benchmarked Qwen 3 8B, Llama 3.1 8B, and Qwen 2.5 7B running locally on an M4 Mac mini against GPT-4o. 19–22 tok/s, unlimited, private. Here are the numbers.