The sizing rule first, because it makes everything else make sense: at 4-bit, a model needs about a gigabyte of memory per billion parameters, plus a few gigabytes of working room. 4-bit means the model's numbers have been rounded down to save space. It costs some quality, but a bigger model rounded down usually beats a smaller one kept precise in the same amount of memory.
The one that surprised me is Qwen3-Coder-Next. It's an 80 billion parameter model, which sounds like data-centre territory, but only 3 billion of those are awake for any given word - the rest sit idle until a word needs their speciality. So it loads like a big model and runs like a small one. You still need about 46 GB of memory for it, so it's one for a well-specced Mac or a pair of graphics cards rather than a laptop.
Down the range
On a single 24 GB card, Qwen3-Coder 30B is I think the best value per gigabyte - the same trick, smaller. Devstral Small (24B) is the pick if you want the model driving your editor across many files. And gpt-oss-20b, OpenAI's open one, runs in 16 GB of ordinary RAM - slowly on the CPU alone, decently on a modest card. On 8 GB, Qwen3 8B is the realistic ceiling. All of these are free to download and openly licensed.
One caveat: most of the benchmark numbers for these are run by the vendors themselves, so I'd treat them as a shortlist, not a ranking.
Easiest way in is Ollama or LM Studio for running the model, plus Continue or Cline in your editor.
Verdict
Worth being honest about the gap, though. Most computers people actually own have 8 or 16 GB of memory and no graphics card worth mentioning, so most of this list is out of reach. And even where a model fits, fitting isn't the same as being usable - on ordinary hardware you watch the words arrive one at a time.
Apple are betting that changes, and that people would rather own their AI than rent it. The new desktop Macs are designed around running models locally: the split between the M5 and M6 chips is really a split in how big a model you can keep loaded, and how comfortably an agent can sit there working all day. The price ladder is asking how much of that compute you think is worth owning. For most people I suspect the honest answer is still none - but I'll be trying the 30B one on the machine under the desk.