What changed
On 2026-07-01 GitHub made Kimi K2.7 Code generally available in Copilot, the first open-weight model offered as a selectable option in the model picker 1. It is beginning to roll out to Copilot Pro, Pro+, and Max plans, hosted by GitHub on Microsoft Azure so you stand up no inference of your own, and billed at provider list pricing under usage-based billing 1. The Stack, reporting independently the next day, framed it as Microsoft opening Copilot to open-weight models for the first time and called Kimi the cheapest non-lightweight option outside GitHub’s fine-tuned Raptor mini 2.
The open-weight bridge
Here is the part the release notes do not say out loud: the copy Copilot hosts on Azure and the copy you can download are the same weights. Kimi K2.7 Code is Moonshot AI’s coding model, a Mixture-of-Experts with 1T total parameters and 32B active and a 256K context window, published on Hugging Face under a Modified MIT License 3. That collapses a decision AI engineers usually agonize over. You can prototype against GitHub’s hosted copy today, then, if data governance or cost pushes you to self-host, stand up the identical weights without switching models or re-baselining behavior. The card’s own numbers are base-model figures, not Copilot scores: Kimi Code Bench v2 rises from 50.9 to 62.0 between K2.6 and K2.7 Code, with roughly 30% fewer thinking tokens than K2.6 3.
| Copilot plan | Kimi K2.7 Code |
|---|---|
| Pro, Pro+, Max | rolling out 1 |
| Business, Enterprise | off by default, admin must enable 1 |
Impact on your team
If you are on Copilot Business or Enterprise, nobody can pick Kimi until an admin flips it on: it is off by default and a plan administrator has to enable the Kimi K2.7 Code policy in Copilot settings 1. That is the first concrete action, before any evaluation. For teams eyeing self-hosting for data governance, the practitioner call is to prototype against the Copilot-hosted copy now and only stand up your own inference once the traces justify the operational cost, since the weights are identical 13. And watch the billing: a model that reads cheap per token can still run expensive per developer at volume under usage-based credits, so meter it before you make it the team default 12.