AINews
Kimi K2.7 Code is the first open-weight model you can select in GitHub Copilot, hosted on Azure and billed by the token.
For all the model choice GitHub has bolted onto Copilot over the past two years, one category was conspicuously missing from the picker: open weights. Every selectable model came from a closed frontier lab. That changed with the general availability of Kimi K2.7 Code, Moonshot AI's open-weight coding model, which is now the first open-weight option you can click on inside Copilot. GitHub is hosting it on Microsoft Azure and billing it under usage-based pricing.
That sentence carries more weight than a routine changelog entry. Microsoft, which owns GitHub and is Copilot's cloud, is now serving a Chinese open-weight model as a first-class citizen alongside the closed models it has historically leaned on. The interesting part isn't that Kimi exists. It's that an open-weight model cleared GitHub's internal quality bar for a production coding surface, and that GitHub decided the cheaper option was worth putting one click away from millions of developers.
What actually shipped #
Kimi K2.7 Code is rolling out gradually to Copilot Pro, Pro+, and Max plans, selectable from the model picker across an unusually wide set of surfaces: VS Code (1.127.0 or later), Visual Studio (17.14.6 or later), JetBrains (1.9.1-251 or later), Xcode, Eclipse, the Copilot CLI, the cloud agent, github.com, and GitHub Mobile. In other words, this isn't a VS Code-only experiment. It's positioned as a general model you can reach from wherever you already work.
The technical envelope matters here. Model registries list Kimi K2.7 Code with a 262,144-token context window, reasoning enabled, and text plus image input. That's a large window by coding-model standards, enough to hold a substantial slice of a repository, a long diff, and your instructions without aggressive truncation. The image input is worth flagging, though as Kimi's own documentation notes, multimodal only works end to end if the editor extension actually passes attachments through correctly, so treat vision support as provider-dependent until you've verified it in your setup.
The lineage is relevant context. Moonshot AI's K2 family is a large mixture-of-experts line released under open weights, which is precisely why you're seeing K2.7 Code show up not just in Copilot but across a long list of inference providers: Cloudflare Workers AI, Fireworks, Together, Hugging Face, Vercel's AI Gateway, and Moonshot's own API all serve it. That's the quiet strategic point. Because the weights are open, GitHub is one host among many, not a gatekeeper.
The economics are the pitch #
GitHub's framing is explicit: this is your "lower-cost option." The pricing bears that out. Provider list pricing for K2.7 Code, as reflected in model registries, lands around $0.95 per million input tokens and $4 per million output, with cached input reads near $0.19 per million.
xychart-beta
title "Kimi K2.7 Code list pricing (USD per million tokens)"
x-axis [Input, Output, "Cache read"]
y-axis "USD / M tokens" 0 --> 4
bar [0.95, 4, 0.19]
Those numbers sit well below what the premium closed frontier models cost per token, and that gap compounds fast in agentic coding, where a single task can burn through hundreds of thousands of tokens across tool calls, file reads, and retries. If you're running Copilot's cloud agent or CLI in a loop, the per-token rate stops being an abstraction and starts being your monthly bill. The sub-twenty-cent cached-read price is especially useful for iterative workflows that re-send the same large context repeatedly, which is most of them.
One caveat: GitHub says the model is billed at "provider list pricing under usage-based billing," so what actually hits your Copilot AI credits may differ from the raw registry figures above, and how each request converts into credits depends on the feature and plan. Treat the numbers as the shape of the cost, not a promise. The direction is clear regardless. Open weights plus commodity inference means downward price pressure, and Copilot is now exposing that pressure directly in the picker.
What this means for how you work #
The practical move is simple. On a Pro, Pro+, or Max plan, open the model picker and select Kimi K2.7 Code. Reach for it on the cost-sensitive, high-volume work: bulk refactors, test generation, boilerplate, first-pass implementations, and long agent runs where you'd rather not meter a premium model. Keep a frontier model on the harder reasoning tasks and switch back when Kimi stalls. The whole value of a picker is that you're not committing to one model for everything, and GitHub's own tooling is leaning into this with auto model selection that routes by task.
Enterprises should read the fine print before celebrating. Kimi K2.7 Code is off by default for Copilot Business and Copilot Enterprise. A plan administrator has to explicitly enable the Kimi policy in Copilot settings, and if it's left off the model simply doesn't appear for anyone in the org. GitHub is unusually direct about why, recommending admins review open-weight models against their own security, compliance, and data-governance requirements first. That's the right posture. "Open-weight" describes the license, not your data path.
And the data path is where the trust boundary shifts. When you pick a default Copilot model, your prompts and responses flow through GitHub's service and its content filters. Selecting a Copilot-hosted Kimi keeps you inside that same Azure-hosted boundary, which is the safe way to consume this model. But because the weights are open and available from a dozen providers, teams may be tempted to wire Kimi into editors through other gateways or self-hosting. The moment you do that, GitHub's filtering and governance no longer apply, and you own the compliance story end to end. Know which door you're walking through.
The bigger signal #
Strip away the changelog packaging and this is a milestone worth marking. Open-weight coding models have spent the last year closing the gap on the closed labs, and their arrival in the default enterprise coding tool is the clearest evidence yet that the gap is small enough to matter commercially. It gives developers real leverage: a cheaper, capable model one click away, and the freedom (for those who want it) to run the same weights elsewhere.
It won't dethrone the frontier models for the gnarliest reasoning work overnight, and the gradual rollout plus enterprise opt-out means adoption will be measured, not instant. But the picker itself is the story. Once an open-weight model is a legitimate, selectable default in Copilot, the question every team starts asking is which tasks actually justify paying premium-model prices. That's a healthy question, and it's now built into the tool.
Sources & further reading #
Kimi K2.7 Code is generally available in GitHub Copilot— github.blog - Kimi AI for GitHub Copilot Chat: How to Use Kimi in VS Code - Kimi— kimi-ai.chat - GitHub Copilot · Plans & pricing— github.com - Kimi K2.7 Code · Models · Pi— pi.dev
Mariana Souza· Senior Editor
Mariana covers the fast-moving world of machine learning and generative AI, with a particular focus on how these technologies are reshaping development workflows. When she isn't stress-testing the latest foundation models, she's usually at a local hackathon.
Discussion 2 #
need to try kimi k2.7 code asap
@zhilakai, hold up, let's not get ahead of ourselves - we still need to see some rigorous evals of kimi k2.7 code, not just a press release 📝