enable Kimi K3 in GitHub Copilot: let Your Team Now

A 2.8-trillion-parameter open-weight model just landed in the model picker next to your chat box. That alone is worth a moment — Kimi K3 from Moonshot AI is the strongest open-weight coding model shipping in Copilot right now, and the only one with a 1-million-token context window at this parameter count. Before anyone on your team pings you asking "is this on yet?", here's what to know and what to do.

If you're on a Business or Enterprise plan, it's off by default. Your admin has to flip the switch first. That's the operational reality behind the launch — and the reason a few teams will hit "model not found" today even though the model is technically live.

What K3 actually is

Kimi K3 is Moonshot AI's open-weight flagship. Two numbers carry the story: 2.8 trillion parameters and a 1-million-token context window. Both are large enough to put K3 in a tier most open-weight models can't reach.

For developers, that translates into two concrete things. First, K3 can ingest a substantial chunk of a real codebase in a single prompt — close to a mid-size repo's worth of context without aggressive trimming. Second, on long-running agentic coding tasks, K3 stays coherent in ways smaller models don't.

GitHub shipped it to Copilot on August 6. The launch itself is unremarkable — a model appearing in a dropdown — but the model underneath is not.

Enabling K3 as an admin

For Business and Enterprise plans, Kimi K3 is gated. It's disabled until an org admin enables it. GitHub's own guidance: review the model against your security, compliance, and data-governance requirements before flipping the switch.

The flow is straightforward:

# 1. Sign in to GitHub as an organization owner
# 2. Settings → Copilot → Models
# 3. Locate "Kimi K3" in the model list
# 4. Toggle the enable switch
# 5. Save

Once enabled, K3 appears in the picker for everyone in the org. Members can then select it themselves from the Copilot chat panel.

Before you click enable

This is the part most teams skip and then regret in postmortem. Three questions worth answering before you toggle K3 on:

1. Where does inference run?  -> Moonshot released the weights, but GitHub
                                  hosts the serving. Your prompts leave your
                                  environment to a third-party endpoint.

2. What does your DPA cover?  -> Confirm the path is covered for the data
                                  classes you care about (PII, source code,
                                  internal docs). Most enterprise DPAs already
                                  have a 15-minute version of this conversation
                                  queued up.

3. Read the model card.       -> Open-weight ≠ audited. The weights are public,
                                  but behavior, jailbreak resistance, and
                                  refusal patterns are not certified the way
                                  a SOC2'd proprietary model is.

If your security team has already signed off on Copilot with Fable 5 or GPT-5.6, the conversation is shorter. If they haven't, K3 is a good reason to have it now.

Switching to K3 as an individual

On Pro, Pro+, and Max plans, there's no admin step. K3 shows up in your dropdown once the rollout reaches your account. GitHub is shipping gradually, so not everyone has it today.

To switch:

1. Open the Copilot chat panel in your editor.
2. Click the model name at the bottom of the panel.
3. Select Kimi K3 from the dropdown.

That's it. Same picker, same flow you already use to swap between GPT and Claude. K3 is available across VS Code, Visual Studio, the JetBrains family, Xcode, Eclipse, the Copilot CLI, github.com, and the mobile clients. One model, every surface.

If K3 doesn't appear yet, that's the rollout, not a bug. Check the GitHub status page or wait a day.

Where K3 wins, where it trails

K3 is the best open-weight coding model shipping right now. It is not the best coding model, full stop — that distinction still belongs to Claude Fable 5 and GPT-5.6 Sol. Knowing the difference matters before you commit a team to it.

Where K3 wins:

• Frontend Code Arena leaderboard: K3 scores 1,679 vs Fable 5's 1,631. Top of the board for frontend work.
• Vals AI Index for real-world tasks: second place.
• SWE Marathon (long-running agent tasks): K3 42.0 vs Fable 5's 35.0. Sustained agentic coding is K3's strongest pitch.

Where it trails:

• SWE-bench Verified: K3 93.4% vs Claude Opus 5's 97% and Fable 5's 95%. Close, but behind on single-shot bug fixing.
• On complex terminal engineering tasks, K3 still trails the proprietary leaders.

The honest summary: if your work is frontend-heavy or your agent runs long enough to lose smaller models, K3 is a serious competitor. If you need the last few percentage points on classic SWE-bench benchmarks, Fable 5 and Opus 5 are still ahead. K3 doesn't replace those — it gives you an open-weight alternative that beats them on the workloads it targets.

What to put in your picker now

A practical default for most teams right now:

- Frontend work, long agent loops  -> K3
- Single-shot bug fixes, tricky   -> Claude Fable 5
  refactors, terminal-heavy work
- Catch-all / "just answer this"  -> GPT-5.6 Sol

Run that for two weeks and you'll have real signal on where K3 earns its slot. Don't pre-commit your whole org to one model on day one — the benchmarks are directional, not definitive, and your codebase's quirks will surface things the leaderboards don't.

Troubleshooting: K3 not in the picker

Three checks, in order:

# 1. Are you on a plan that has it?
#    Business / Enterprise -> admin must enable
#    Pro / Pro+ / Max      -> rollout-dependent, may not have hit your account

# 2. Is the rollout live for you?
#    github.com status page lists Copilot rollout progress.
#    "Model not found" usually = rollout hasn't reached your account.

# 3. Did you restart the editor?
#    The model list sometimes needs an IDE reload after org changes.

If all three check out and K3 still isn't there, your admin hasn't enabled it (Business/Enterprise) or you're behind the rollout wave (individual plans). GitHub Support is the next step, but realistically — wait 24 hours and check again.

The August 6 hiccup

Worth flagging because admins will hear about it. On August 6, a 10-hour GitHub Actions outage throttled webhooks and broke Copilot code review. The K3 rollout was briefly paused during that window. The outage has since been resolved and the rollout is back on track.

If your team saw Copilot code review go dark on August 6, that was the Actions incident, not K3. If you're writing a postmortem, pull the GitHub status archive entry — it's resolved.

The part that doesn't change when the model does

Here's the bit that doesn't fit in any release note. Models churn. K3 is the best open-weight coding model right now — until the next open-weight model ships, and the next one after that. The proprietary leaders will keep swapping too. Your model picker will look different in six months than it does today.

The parts that should not churn with the model: your auth flows, your deployment pipeline, the components your team actually clicks on, and the templates your team starts new projects from. A frontend repo is a frontend repo whether the suggestions come from K3, Fable 5, or whatever ships next quarter. Pick a model because it's the best model for the task at hand. Build the rest of your stack on primitives that survive the swap.

Pick K3 today. Keep the rest of the stack stable when the next one lands.