Local AI on an Intel Mac: Tested and Analysed

A ten-part account of building a free local coding agent on a 2019 Intel Mac, from hardware checks and failed downloads to Radeon acceleration, benchmarks, tool-use tests and a full retest after a reboot.

Episodes

Published in these episodes

Cover: a 2019 16-inch MacBook Pro keyboard in a circle beside the Ollama and Qwen marks, labelled MacBook Pro 2019, Core i9, 32 GB

AI & Automation

Can a 2019 Intel MacBook Pro Run a Free Local AI Coding Assistant?

Before installing AI models and coding agents, check what an older Intel Mac can handle, and which GPU claims still need proof on your own machine.

Cover: the Ollama and Qwen marks beside the Ollama prompt box, Kilo Code’s Review First card and a Hello World page, with the command ollama run qwen3.5:4b

AI & Automation

Install Ollama and Qwen

Kilo Code, Twinny and Claude Code against a local Ollama server on an Intel Mac: a working API, a real generated webpage, and the limits small models hit when they drive coding agents.

Cover: the Ollama, Qwen and Python marks over the qwen3.5 4B and 9B rows of Ollama’s model table, and a download bar that breaks at 48% and resumes to 100%

AI & Automation

Recover an Interrupted Ollama Model Download: Qwen 9B, unexpected EOF, and a Better Resume Strategy

A practical recovery guide for interrupted Qwen downloads, from unexpected EOF failures at 48% and 91% to safe retries, resumable transfers, integrity checks, and final Ollama verification.

Cover: the chain from llama.cpp through Vulkan and MoltenVK to the Radeon Pro 5500M, beside the llama.cpp banner from llama-cli

AI & Automation

Run Local LLMs on an Intel Mac's AMD Radeon GPU with Vulkan, MoltenVK, and llama.cpp

Ollama ran Qwen on the CPU of a 2019 Intel MacBook Pro. A Vulkan-enabled llama.cpp build finally put the machine's Radeon Pro 5500M to work, with verified device detection and model offloading.

Cover: the Qwen mark and a graphics card beside the prompt processing and generation charts for Qwen 3.5 4B and 9B on the Radeon

AI & Automation

AMD Radeon Pro 5500M LLM Benchmarks: Qwen 4B vs. 9B with llama.cpp and Vulkan

Measured Qwen3.5 4B and 9B performance on a 2019 Intel MacBook Pro after Vulkan and MoltenVK enabled its Radeon Pro 5500M, with commands, offloading evidence, and honest limits.

Cover: the llama.cpp and bash marks and the word Vulkan over a pattern of folders, beside four steps of the move: Copy, Rebuild, Launch and Review

AI & Automation

Make a Local AI Installation Permanent: Organise llama.cpp, Rebuild Vulkan, and Reuse Ollama Models

Move a working llama.cpp experiment out of Downloads, rebuild Vulkan cleanly, reuse existing Ollama Qwen weights, and create a permanent launch workflow on an Intel Mac.

Cover: OpenCode taking the request to find the homepage button, the built-in tools menu of llama.cpp’s web chat, the OpenCode, llama.cpp and Qwen marks and an offline sign

AI & Automation

Build an Offline Coding Agent with OpenCode and llama.cpp on an Intel Mac

Connect OpenCode 2.1.2 to a Radeon-backed local Qwen 9B server, configure approval-driven tools, recover from an 8K context failure, and verify project file access at 32K.

Cover: four steps, read, write, approve and run, over the agent’s hello.js code and its output 5, with the OpenCode, Node.js and JavaScript marks

AI & Automation

Testing a Local AI Coding Agent: Read, Write, Approve, Execute

A real OpenCode test in which Qwen 9B read project configuration, created hello.js after approval, executed it through Node.js, and correctly reported the output 5.

Cover: a timer reading 8:19 over the agent’s Fibonacci output ending in runs of zeros and its letters turning to asterisks, with the Qwen and OpenCode marks

AI & Automation

Eight Minutes for Wrong Answers: How to Evaluate a Small Local AI Coding Agent

A Qwen 9B agent completed a Fibonacci task in 8 minutes 19 seconds, but ordinary JavaScript Number arithmetic corrupted large integers. The failure shows how to test coding agents for correctness.

Cover: an 8 GB ceiling over the Qwen mark and the names Gemma 4 and Granite 4.2, beside memory plans where two settings pass the limit and flash attention brings them under

AI & Automation

After the Reboot: Five Local Models, One Context Setting and an 8 GB Ceiling

A full restart, then every command checked again: Qwen 3.5, Granite 4.2, Ministral 3 and Gemma 4 on a 2019 MacBook Pro's Radeon Pro 5500M. The command layer came back intact; the 8 GB card, one model's packaging and the Mac's thermal cap decided the rest.

Ready when you are

Tell us what you need and a written quote comes back within a working day.

or connect with us via :

WhatsApp Email