GitHub Copilot Added Kimi K3: How to Enable It (and Why GitHub Paused the Rollout on Day One)

On August 6, 2026, GitHub shipped Kimi K3 — Moonshot AI's open-weight model — as a selectable model inside GitHub Copilot. Then, the same day, GitHub added an editor's note to its own changelog post saying it had paused the rollout to mitigate an unrelated GitHub Actions incident. If you went looking for Kimi K3 in your model picker that afternoon and didn't find it, you weren't misconfigured — you were just on the wrong side of a rollout that GitHub itself interrupted mid-flight.

That timing aside, this is a real, practically useful addition: Kimi K3 is the first genuinely frontier-class open-weight model available inside Copilot's agentic coding flow, at a fraction of the per-token cost of Copilot's frontier closed models. This post covers what it actually is, exactly where the enablement toggle lives (it's not where most admins expect), what it costs, and how to tell whether you're looking at a policy problem or just the rollout pause.

What Kimi K3 actually is

Kimi K3 is Moonshot AI's latest open-weight large language model, hosted for Copilot through Fireworks AI rather than run directly by GitHub. GitHub's own changelog describes it as showing "frontier-level abilities on agentic coding with highly cost-effective pricing" — the second half of that claim is the interesting part. Unlike GitHub's own first-party models, Kimi K3 is priced as a usage-based model with published per-token rates rather than folded into a flat premium-request allowance, which matters once you start running it inside Copilot's agent mode across a large codebase.

It's rolling out across every Copilot surface: VS Code, Visual Studio, the Copilot CLI, github.com's web chat, and the Copilot mobile apps. It's available on Copilot Pro, Pro+, Max, Business, and Enterprise plans — but "available" and "enabled" are different things for Business and Enterprise customers, which is where most of the confusion starts.

Why this needed an admin toggle at all

For individual Copilot Pro/Pro+/Max subscribers, there's nothing to configure — Kimi K3 just appears as a new entry in the model picker once the rollout reaches your account. For Copilot Business and Copilot Enterprise, GitHub shipped it off by default. An org or enterprise admin has to explicitly turn it on before anyone in that org can select it, same pattern GitHub has used for every new third-party-hosted model added to Copilot this year (Kimi K2.7 in July worked the same way).

The reasoning is straightforward: usage-based models bill differently than the flat-allowance models bundled with a Copilot seat, and GitHub doesn't want an org to accidentally rack up per-token charges on a model nobody in IT signed off on.

Step-by-step: enabling Kimi K3 for an organization

1. Go to your organization's page on github.com and open Settings.
2. In the left sidebar, under Code, planning, and automation, click Copilot.
3. Click Models (not Policies — that page controls privacy and feature toggles like Copilot Chat or code referencing, not individual third-party models).
4. Find Kimi K3 in the model list and set it to Enabled. If your enterprise owner has already set an enterprise-wide policy for this model (Enabled or Disabled for everyone), the org-level toggle will be greyed out and show the inherited state instead — you can't override an enterprise-wide "Disabled."

For enterprise-level control instead of per-org:

1. Go to your enterprise account and open the AI controls tab.
2. Locate the model policy for Kimi K3.
3. Choose Enabled for everyone, Disabled for everyone, or Optional (lets individual organizations under the enterprise decide for themselves).

If you only need specific organizations to have access rather than the whole enterprise, use targeted model rules instead of a blanket enterprise policy — available on Business and Enterprise plans — which let you allow a model for selected organizations without touching the enterprise default.

1Organization settings
2  └── Copilot
3        └── Models → find "Kimi K3" → Enabled
4
5Enterprise account
6  └── AI controls
7        └── Model policies → Kimi K3 → Enabled / Disabled / Optional

Once enabled, developers pick it the same way they'd pick any other model — from the model dropdown in Copilot Chat, the CLI's --model flag, or the agent-mode model selector — no per-user opt-in step beyond the org/enterprise policy.

What it costs

Kimi K3 in Copilot is billed usage-based, not against your flat premium-request allowance:

That's meaningfully cheaper than Copilot's frontier closed models on output tokens, which is exactly where agentic coding sessions rack up cost — long tool-call loops generate a lot of output relative to input. If your team runs Copilot's agent mode against large repos for extended sessions (refactors, multi-file migrations, test generation sweeps), this pricing is the actual reason to care about this release, not the model announcement itself.

Best practices

• Enable it at the org level first, not enterprise-wide, if you're only testing. Targeted model rules let you validate cost and quality with one team before flipping it on for every organization under your enterprise.
• Watch cached-input pricing if you're doing repeated agent-mode runs against the same repo context. At $0.30 per 1M cached tokens versus $3 for fresh input, repo-aware caching matters more here than with flat-allowance models where you never see the per-token breakdown.
• Treat "off by default" as intentional, not a bug. Every new Copilot model this year has shipped this way for Business/Enterprise. Build the org-level enablement step into your standard "new model available" checklist instead of treating it as a one-off surprise each time.

Common mistakes to avoid

• Looking for it under Policies instead of Models. The Policies page controls features (Chat, code referencing, public code matching); individual model availability is a separate Models page.
• Assuming it's missing because of misconfiguration when it's actually the rollout pause. GitHub explicitly noted it paused Kimi K3's rollout on August 6 to mitigate an unrelated Actions incident, then resumed it. If the model doesn't appear even after you've confirmed the policy is Enabled, check GitHub's changelog post for the current rollout status before assuming your settings are wrong.
• Trying to override an enterprise "Disabled" policy at the org level. If your enterprise owner has locked the model off, the org Models page will show that inherited state and won't let you flip it — the fix is a conversation with the enterprise owner, not more clicking around org settings.

Troubleshooting

The Models page doesn't show Kimi K3 at all. The rollout is gradual across accounts; a temporary absence from the list (versus it appearing but greyed out) usually means your organization hasn't been reached yet, or you're caught in the pause window GitHub described in its changelog note.

Toggle is enabled but developers still don't see it in the model picker. Model picker availability can lag policy changes by a few minutes to a few hours as it propagates. If it's been over a day, verify there isn't a conflicting enterprise-level policy overriding your org setting.

Usage-based billing charges appear unexpectedly on your Copilot invoice. Confirm which models on your Models page are set to usage-based versus included-in-seat. Kimi K3, like other third-party-hosted models, bills per token regardless of your Copilot plan tier.

FAQ

Is Kimi K3 free with my Copilot subscription? No — it's billed usage-based per token on top of your Copilot seat, unlike GitHub's own bundled models which draw from your premium-request allowance.

Do individual Copilot Pro users need to enable anything? No. The admin toggle only applies to Business and Enterprise plans. Pro, Pro+, and Max subscribers see it appear directly in their model picker as the rollout reaches their account.

Is Kimi K3 actually good for agentic coding, or is this just a pricing story? Both are true. Moonshot AI has positioned K3 specifically around agentic coding benchmarks, and independent testing since its open-weight release has generally shown it competitive with — not universally better than — established frontier models. The pricing is what makes it worth testing rather than defaulting to a closed model out of habit.

Does enabling Kimi K3 affect GitHub Copilot's data retention or training policies? GitHub's changelog doesn't call out different data handling for Kimi K3 versus other Copilot models — it runs through GitHub's existing Copilot infrastructure via Fireworks AI, not a direct connection to Moonshot AI. If your org has strict data-handling requirements, verify current terms in GitHub's Copilot documentation before enabling any new model broadly.

What happened with the GitHub Actions incident that paused the rollout? GitHub's changelog only says the rollout was paused "to mitigate an incident with GitHub Actions" without detailing the incident itself — treat that as unrelated to Kimi K3's quality or safety and specific to whatever Actions infrastructure issue was happening that day.

Key takeaways

Further Reading

• GitHub Models Retires July 30, 2026: How to Migrate Your Code and CI Workflows
• GitHub Actions Now Holds Suspicious Workflow Runs Before They Touch Your Secrets
• MCP Goes Stateless on July 28, 2026: What Breaks and How to Migrate Your Server
• Kimi K3 is now available in GitHub Copilot (official changelog)
• GitHub Copilot policies for enterprises and organizations (official docs)