Which open coding models can you run on your own computer?

An open-weight model is one whose trained files you can download and run on your own hardware, with no account, no per-token bill and no code leaving your machine. For coding, that matters if you work on private code, travel without a reliable connection, or simply want to learn how these systems behave up close.

The catch is size. The models at the top of the coding leaderboards are hosted only, and the largest open releases need server racks. But a middle class of models, around 27 to 30 billion parameters (the numbers a model learns during training), now fits on a single good home machine and handles real coding tasks.

The two best picks this fall

SWE-bench Pro is a test made of real bug fixes and features in open-source projects. Both figures come from the makers, and Qwen's card shows Qwen3.8-27B ahead of Glimmer on several coding tests; Qwen even lists it above an older Claude model, Opus 4.6, on that test. Treat those as promising signs, not proof, and check them on your own code.

Apache 2.0 is a permissive licence: you may use either model at work, change it and ship products with it, as long as you keep the licence notice.

Starting one with Ollama

Ollama is a free tool that downloads models and runs them on your computer behind a simple command. Once it is installed, two commands get you a working coding setup:

# Download (about 18 GB) and chat with Qwen3.8 27B
ollama run qwen3.8

# Or start a coding agent that uses the local model
ollama launch claude --model qwen3.8

The second line, shown on Ollama's model page, starts Claude Code with the local model in place of Anthropic's servers; the same page lists OpenCode as another option. Swap in muse-glimmer to try Meta's model instead. If you prefer llama.cpp for Glimmer, Meta's card notes that you need build b10353 or newer.

Then give it a task you can check:

Read src/cart.py and tests/test_cart.py. Add a function that applies a percentage discount to the cart total, rounding to two decimals. Write three tests for it, including a 0% and a 100% discount, run them, and show me the diff.

Local models are slower than hosted ones and make more mistakes on long, multi-step work, so start with focused tasks: one function, one test file, one bug. Both models think before answering by default; Qwen lets you switch that off per request when you want quick replies.

What hardware do you really need?

The rule is simple: the model file has to fit in your graphics card's memory, or in the shared unified memory of an Apple-silicon Mac, with room left over for the conversation. That leftover room holds the context, the code and messages the model is reading, and long contexts take a lot of it.

The 4-bit builds store each parameter in about half a byte instead of two, which is how a 27-billion-parameter model shrinks to about 18 GB. Meta says its 4-bit Glimmer builds lose little or no accuracy on agent tasks; for Qwen, judge for yourself on your own tasks.

What will not fit at home are the giants. Meta's Muse Spark Open, for example, has 816 billion parameters, which even at half a byte each is roughly 400 GB of weights, a job for data-centre graphics cards. For those, a hosted API is the practical route.

If you are buying hardware for this, memory matters more than raw speed. A card or Mac with more memory runs bigger models and longer contexts; a faster chip with too little memory simply cannot load them.