GitHub Ships K3 to Every Copilot Tier

GitHub shipped K3 to every Copilot tier on August 13, per its changelog.

Developers who waited for community-built local runtimes to catch up got the model through Copilot first, weeks before those runtimes could run it at all. GitHub's hosted rollout became the fast path to K3, not the open-weight download.

K3's weights went public in late July. It's the largest open-weight model released so far. Local inference engines still took weeks to add support. Weight availability wasn't the bottleneck holding K3 back. Runtime support was, and GitHub's own integration got there first.

Sub-1B models make up 83% of local downloads, and that's where quantization allocation research is actually paying off right now. K3 sits at the frontier end, far from where those local gains are landing.

A model that's open in name but takes weeks before an inference engine can run it locally is closed in practice for anyone without a hosted subscription. Watch how long the next big open-weight release takes to run outside a platform like Copilot. That gap, not the license, is becoming the real measure of how open a model is.