Kimi K3 Is Here: What Moonshot's 2.8T Model Means for Your Studio

On July 16, Moonshot AI released Kimi K3, and it's the biggest "open" model announcement of the year in the most literal sense: roughly 2.8 trillion parameters (Mixture of Experts), a 1 million token context window, image input, and launch benchmarks that claim wins over several Western flagships. The weights are promised as a public release by July 27, 2026.

Here's what that means in practice, and what it means for LU Labs Cloud.

What K3 is

K3 is Moonshot's frontier swing: a new MoE architecture aimed at long-horizon coding and agent work. Two variants shipped at launch, K3 Max (chat and agents) and K3 Swarm Max (large-scale parallel agent workloads). The headline features are the 1M token context (entire codebases or book-length documents in one window) and the benchmark claims, which put K3 ahead of Claude Opus 4.8 max and GPT-5.5 high on Moonshot's own tests.

Two honest notes. First, those are vendor benchmarks; independent verification needs the weights, and the weights aren't out yet. Second, K3 currently runs only on Moonshot's servers. Even OpenRouter forwards every request straight to Moonshot. There is no privacy choice to make about K3 today, because there's only one place it lives.

Why we're not adding it (yet)

LU Labs Cloud runs on a simple rule: open-weight models on infrastructure we choose, at flat monthly prices, with your data staying between you and the render. K3 doesn't fit that rule yet:

• No weights, no independent hosting. Until July 27 (Moonshot's promised date), nobody but Moonshot can serve K3.
• Routing every prompt to the model's maker isn't the deal we offer. When neutral providers can host K3, that changes.
• Launch pricing is steep ($3 per million input tokens, $15 per million output, reasoning billed as output at max effort only). We do the margin math before anything enters the catalog, so flat plans stay flat.

The moment K3 is hostable outside Moonshot and the numbers make sense, it becomes a catalog candidate like any other open model. Our catalog is server-driven, so when a model is added it appears in the web app and the desktop app immediately, no update required.

The smart pick today: Kimi K2.6

Here's the part launch-day coverage skips: Moonshot's Kimi K2.6 is already one of the strongest open models for agentic and coding work, its weights are public (Modified MIT), it's served by independent providers, and it's in the LU Labs Cloud catalog right now, alongside 30+ other open models on your plan.

For most sessions, K2.6 is the pragmatic choice: proven over months where K3 is a week-old claim, roughly a third of the input price, and available with provider choice. A long agent session that costs cents on K2.6 can cost dollars on K3.

The takeaway

• Today: use Kimi K2.6 in LU Labs Cloud for agentic coding and long-document work. It's live in the model picker.
• July 27: the K3 weights drop. Independent benchmarks and hosting follow.
• After that: if the margins and quality hold up, K3 joins the catalog, and you'll see it appear in your studio without lifting a finger.

Not on Cloud yet? See the plans. Every tier includes the full open-model catalog: chat, code, image, and video in one studio.