Skip to content
NewsModel release

Moonshot AI's Kimi K3, a 2.8-trillion-parameter open-weight model, joins GitHub Copilot at $3 per million input tokens

· by Pondero Newsdesk

The short version

Kimi K3 reached general availability in GitHub Copilot on August 6, 2026, giving subscribers access to the largest open-weight model deployed in a mainstream coding tool, hosted on Fireworks AI under usage-based billing.

Moonshot AI's Kimi K3, a 2.8-trillion-parameter open-weight model, joins GitHub Copilot at $3 per million input tokens

GitHub Copilot added Kimi K3 to its model picker on August 6, 2026. That makes it the first time a model in the multi-trillion-parameter class has been accessible through a mainstream developer subscription rather than a self-hosted inference stack. Priced at $3 per million input tokens and $15 per million output tokens under Copilot's usage-based billing, the model is available to subscribers on Pro, Pro+, Max, Business, and Enterprise plans.

What

Kimi K3 became generally available in GitHub Copilot on August 6, 2026, hosted by GitHub on Fireworks AI, per the GitHub Changelog. Subscribers can select it through the model picker in VS Code, Visual Studio, JetBrains, Xcode, Eclipse, the Copilot CLI, Copilot cloud agent, Copilot app, github.com, and GitHub Mobile on iOS and Android. Cached input tokens are billed at $0.30 per million.

For Business and Enterprise customers the model is off by default. Organization administrators must enable the Kimi K3 policy in Copilot settings before anyone in their organization can access it. GitHub's changelog recommends that admins review open-weight models against their organization's security, compliance, and data-governance requirements before enabling them.

Kimi K3 uses a mixture-of-experts architecture with 896 total experts and 16 activated per token, producing 104 billion active parameters out of 2.8 trillion total, per Moonshot AI's technical documentation. The model runs 93 layers, uses two architectural components called Kimi Delta Attention and Attention Residuals, and supports a one-million-token context window. Native multimodal capabilities let it process text, images, and video within the same model. Moonshot AI says the architectural changes produce roughly 2.5 times better scaling efficiency compared to its predecessor Kimi K2, though that figure is a vendor self-assessment.

GitHub briefly paused the August 6 rollout to handle an unrelated GitHub Actions incident, then resumed the same day, per the changelog.

Why it matters

Size alone does not determine how useful a model is on a given coding task, but the parameter count does set a concrete ceiling on what the model can represent. Until now, models at this scale were either proprietary or required self-hosted infrastructure that most teams do not run. Copilot's Fireworks AI hosting removes that barrier for subscribers who already pay for the service.

Pricing puts Kimi K3 in a directly comparable range to Claude Sonnet-class models available in Copilot at similar token rates. Teams optimizing for cost relative to proprietary frontier models now have a new option. Per Moonshot AI's own evaluation suite, Kimi K3 outperforms other tested open-weight models on agentic coding benchmarks while trailing the two strongest proprietary models in their suite, Claude Fable 5 and GPT 5.6 Sol. That positions it as a middle tier: stronger than previous open-weight options in Copilot, not yet matching the closed-source ceiling.

The admin opt-in for Business and Enterprise is a real scheduling consideration. Any developer expecting to pick up Kimi K3 tomorrow on a Business or Enterprise plan should check with their GitHub organization admin first. The policy has to be enabled before the model appears in the picker. That is a one-time action, but it requires coordination that individual Copilot Pro or Pro+ subscribers do not need.

Open-weight status also carries an implication beyond performance. The Kimi K3 weights are available under Moonshot AI's own license for research and deployment outside Copilot. Organizations with unusual governance requirements around model portability or auditability have more options with an open-weight model than a black-box API.

Context and reactions

Moonshot AI released Kimi K3 in late July 2026, describing it as the world's first open 3T-class model. Per the Kimi tech blog, Kimi models have held the upper bound of open-model parameter counts for nine of the twelve months leading up to K3's release. The model extends that run into territory that was previously closed to open-weight research.

GitHub has been expanding Copilot's model roster throughout 2026. Adding Kimi K3 means Copilot now offers models from Chinese, American, and European labs alongside its own infrastructure. Enterprise customers should examine the Fireworks AI routing before enabling Kimi K3 at scale. Data and requests pass through Fireworks AI's infrastructure before Copilot returns a response, which matters for organizations with strict data-residency requirements.

The brief rollout pause on August 6 had no connection to Kimi K3 itself. GitHub resolved the Actions incident and restored the model the same day.

What to watch next

GitHub has not announced plans to extend Kimi K3 access to the free Copilot tier, where the model picker is restricted. Free-tier inclusion would be the single largest expansion of the addressable audience for K3 within Copilot, and any announcement in that direction would signal a shift in how GitHub prices open-weight model access.

Independent evaluations from developers running Kimi K3 against Claude Sonnet and GPT-5.6 Luna in real Copilot workloads will determine whether Moonshot AI's benchmark advantages on agentic coding hold in practice. Those comparisons typically emerge within weeks of a new model landing in a widely used tool.

Sources