# Kimi K3 Is Now in GitHub Copilot: Enable It Before Your Team Asks > Source: > Published: 2026-08-09 01:10:52+00:00 Moonshot AI’s Kimi K3 — a 2.8-trillion-parameter open-weight model with a 1-million-token context window — landed in [GitHub Copilot on August 6](https://github.blog/changelog/2026-08-06-kimi-k3-is-now-available-in-github-copilot/). If you’re on a Business or Enterprise plan, it’s off by default, and your admin needs to enable it before anyone on the team can use it. That’s the first thing to know. Here’s the rest. ## Enabling Kimi K3 (Especially If You’re an Admin) For Copilot Business and Enterprise, Kimi K3 is disabled until an organization admin turns it on. GitHub explicitly recommends reviewing the model against your security, compliance, and data-governance requirements before flipping that switch. Once enabled, team members can select it from the model picker in Copilot settings. On individual plans — Pro, Pro+, and Max — it shows up automatically in the model dropdown once the rollout reaches your account. GitHub is rolling out gradually, so not everyone has it yet. To switch models: open the Copilot chat panel, click the model name at the bottom, and select Kimi K3. It’s available across VS Code, Visual Studio, JetBrains, XCode, Eclipse, Copilot CLI, GitHub.com, and mobile. One note: the rollout was briefly paused on August 6 due to a 10-hour GitHub Actions outage that throttled webhooks and broke Copilot code review. That has since been resolved, and the rollout is back on track. ## What the Benchmarks Actually Say Kimi K3 is the best open-weight coding model available right now. It is not the best coding model, period — that distinction still belongs to Claude Fable 5 and GPT-5.6 Sol. Understanding the difference matters before you decide whether to use it. Where K3 wins: it takes the top spot on the [Frontend Code Arena leaderboard](https://www.vals.ai/benchmarks/swebench) (score: 1,679 vs Fable 5’s 1,631), ranks second on the Vals AI Index for real-world tasks, and beats Fable 5 on long-running agent tasks (SWE Marathon: K3 42.0 vs Fable 5’s 35.0). For frontend work and sustained agentic coding sessions, K3 is a genuine competitor to the proprietary leaders. Where it trails: on SWE-bench Verified, K3 scores 93.4% against Claude Opus 5’s 97% and Fable 5’s 95%. On complex terminal engineering tasks, Fable 5 maintains a clear lead. Moonshot acknowledges this directly — K3 trails Claude Fable 5 and GPT-5.6 Sol in overall composite rankings. The honest framing: if you need the highest success rate on hard backend engineering tasks, K3 isn’t the move. If you do a lot of frontend work or run long multi-step agent loops, it’s worth serious consideration. ## What “Open Weight” Changes Here Every other model in the GitHub Copilot picker is proprietary. Kimi K3 is the exception: the weights are on [Hugging Face under an Apache 2.0 license](https://huggingface.co/moonshotai/Kimi-K3), which means you can self-host and fine-tune commercially. This matters most for three groups. Regulated industries — finance, healthcare, legal — can run K3 on their own infrastructure and keep all data in-house. Teams with specialized needs can fine-tune the weights for domain-specific tasks (legal code review, infrastructure automation) in ways you simply can’t do with Claude or GPT. And at high enough volume, self-hosting sidesteps per-token billing entirely. The catch is real: self-hosting Kimi K3 requires a minimum of 64 GPU accelerators and roughly 1,680 GB of VRAM. That is enterprise infrastructure, not side-project territory. For anyone without that kind of hardware, the [Copilot-hosted version on Fireworks AI](https://northflank.com/blog/what-is-kimi-k3-self-hosting) is the practical path. ## Pricing Kimi K3 runs $3 per million input tokens and $15 per million output tokens under Copilot’s usage-based billing. Cache hits drop to $0.30 per million input tokens — a 90% discount that becomes significant for long-context workflows where you’re repeatedly querying the same large codebase. That’s more expensive than Kimi K2.6 (which ran $0.95/$4 on input/output) and roughly on par with Anthropic’s Claude Sonnet pricing. Moonshot is positioning K3 as a premium model, not a budget alternative. ## Who Should Enable It Now Enable K3 if your team does heavy frontend development, runs long agentic coding sessions, or needs 1-million-token context for large codebases. The benchmarks support it in those specific use cases. Wait if your primary need is the highest raw success rate on difficult backend engineering tasks — Claude Fable 5 still leads there. Also wait if you’re on Business or Enterprise and haven’t completed a compliance review; GitHub’s recommendation to do that first exists for a reason. For regulated industries evaluating whether to self-host: the infrastructure requirements are documented and the Apache 2.0 license supports commercial deployment, but plan for significant hardware investment. Full technical details are on [Moonshot’s tech blog](https://www.kimi.com/blog/kimi-k3).