GitHub Copilot Just Added Its First Open-Weight Model — and It's Chinese
Moonshot AI's Kimi K2.7 Code is now generally available in Copilot's model picker — hosted by GitHub on Azure, so your code need never touch Moonshot's servers. In the week of the Fable 5 compact, the open-weight lane keeps flowing through the export-control chess match.

The takeaway: GitHub quietly crossed a line on Wednesday 1 July: Kimi K2.7 Code — an open-weight model from China's Moonshot AI — is now generally available in GitHub Copilot, and it's the first open-weight model ever offered in Copilot's model picker. In the same week that Washington's grip on who gets American frontier models dominated the news, Microsoft's flagship developer tool started serving a Chinese model — hosted on Microsoft's own cloud — as its budget option.
What GitHub announced
Per GitHub's changelog, dated Wednesday 1 July:
- Kimi K2.7 Code is generally available in Copilot, rolling out gradually to Copilot Pro, Pro+ and Max plans, selectable in the model picker in Visual Studio Code, with expansion to Business and Enterprise "over the coming weeks".
- It's pitched as "a lower-cost option for your coding workflows" — the value proposition is price, not peak capability.
- It's the first open-weight model in the picker. Every other Copilot model option — the GPT line, the Claude line, Gemini, Microsoft's own MAI — is a closed model: the weights stay private, and you reach it only as a service. Kimi K2.7 Code is the first whose weights anyone can download. (Claude Sonnet 5 joined the picker the day before, for scale of how fast that menu now moves.)
- For enterprises it's off by default: Copilot Business and Enterprise administrators must explicitly enable the Kimi K2.7 Code policy before anyone in their organisation can select it.
That last point, plus one hosting detail, does a lot of quiet work.
The hosting detail that matters
GitHub says Kimi K2.7 Code is hosted by GitHub on Microsoft Azure, billed at provider list pricing under usage-based billing. Because the model is open-weight — the weights are downloadable, the shift we unpacked in our open-weight state-of-play piece — Microsoft doesn't need to send a single byte of your code to Moonshot AI's servers. The model runs on Microsoft infrastructure; Moonshot need never see the traffic.
That's what makes this arrangement politically viable in 2026. An American enterprise routing its source code through a Chinese company's API would be a hard sell — to security teams and, increasingly, to regulators. An American company running a Chinese-origin model on American servers is a different proposition entirely. Open weights are what make the difference: the model becomes an artefact you host, not a service you call.
Why this lands harder this particular week
The timing gives the story its edge. This is the week the US-China contest over AI access has been front and centre: Fable 5 returned from its export ban under a negotiated security compact, GPT-5.6 remains gated to a small group of government-blessed partners, and a CNBC analysis put China's open-source GLM-5.2 within about a point of Opus 4.8 on a closely watched agentic benchmark at a fraction of the cost.
The pattern we flagged in our DeepSeek DSpark coverage keeps holding: while US frontier labs compete on capability with increasingly gated access, Chinese labs compete on cost and openness — and openness is proving to be a distribution strategy. Moonshot didn't need to route Copilot's traffic through its own API to get Kimi into the picker. It needed weights good enough, and open enough, that GitHub could pick them up and host them itself. Kimi K2.7 posted eye-catching first-party coding numbers at launch in mid-June; three weeks later it's the budget lane in what GitHub bills as the world's most widely adopted AI developer tool.
The honest caveats
GitHub's own framing is modest: this is a lower-cost option, a gradual rollout, and the company says it will "continue to monitor the model's quality and performance". Nobody — GitHub included — is claiming K2.7 matches the closed frontier models in the same picker. The enterprise off-by-default policy also signals that Microsoft knows some customers will want nothing to do with a Chinese-origin model regardless of where it's hosted; that choice now sits with each company's administrators.
But the precedent is set: the model picker in Microsoft's flagship developer product now includes a downloadable Chinese model as a first-class citizen. However the export-control chess match plays out at the frontier, the open-weight lane keeps flowing straight through it.
That's the story. As ever: On The Wire runs on Anthropic models — one of Kimi's closed-model rivals in that same picker — so read our model coverage with that on the table.
Ask Relay — he reads every question himself and replies personally by email.
