Moonshot AI has launched Kimi K3, a 2.8-trillion-parameter flagship model that pushes Kimi deeper into frontier coding, long-context reasoning, multimodal work, and agentic knowledge tasks.

The headline is scale, but the more important story is access. Moonshot describes Kimi K3 as the first open model in the 3-trillion-parameter class. The model is available now through Kimi.com, Kimi Work, Kimi Code, and the Kimi API, while Moonshot says the full model weights are scheduled for release by July 27, 2026.

That puts Kimi K3 in the same strategic lane as DeepSeek's biggest open-model moments: a Chinese AI lab is trying to make a frontier-class model broadly usable, relatively cheap through an API, and visible on coding and agent benchmarks where developers can compare it against closed U.S. systems.

Official Kimi demo

Kimi's launch material includes official videos showing K3-generated interactive and visual work. This clip from Kimi's own K3 blog shows a browser-based 3D open-world demo generated as part of the model's game-development case studies.

Moonshot also posted an official short launch video on X with the caption "Meet Kimi K3."

What Kimi K3 is

Kimi K3 is Moonshot's new top model for long-horizon coding, end-to-end knowledge work, visual reasoning, and deep reasoning.

The core specs are unusually large for an open model:

  • 2.8 trillion total parameters
  • 1 million-token context window
  • native visual understanding
  • Kimi Delta Attention and Attention Residuals
  • Stable LatentMoE with 16 of 896 experts active per token
  • default max reasoning effort at launch

Moonshot says the architecture gives K3 roughly 2.5x the scaling efficiency of Kimi K2. The model is still sparse at inference time, which matters because a 2.8-trillion-parameter dense model would be wildly expensive to serve. K3's Mixture-of-Experts setup is the practical trick: the model is huge overall, but only a small slice is active for each token.

The API is priced at $0.30 per million tokens for cache-hit input, $3.00 per million tokens for cache-miss input, and $15.00 per million output tokens. That is not bargain-basement pricing, but it is aggressive for a model Moonshot wants compared against frontier proprietary systems.

Why developers are paying attention

Kimi K3 is landing hardest with developers because Moonshot is positioning it as an agentic coding model, not just a chat model with better benchmark numbers.

Moonshot says K3 can sustain long-running engineering tasks, navigate large repositories, coordinate terminal tools, use screenshots and visual feedback, and work across frontend engineering, game development, CAD-like workflows, kernel optimization, compiler work, and research coding.

The company highlights several technical case studies:

  • K3 optimized GPU kernels across tasks involving AttnRes, KDA, and large-head MLA kernels.
  • K3 built MiniTriton, a compact Triton-like compiler stack with a tile-level IR over MLIR.
  • K3 generated playable browser game demos with Three.js and visual feedback loops.
  • K3 designed, optimized, and verified a small chip for a nano model in a 48-hour autonomous run.
  • K3 reproduced a computational astrophysics workflow by reading papers, implementing equations, writing thousands of lines of Python, and producing an interactive dashboard.

Some of those claims still need outside replication, especially the more cinematic case studies. But the pattern is clear: Moonshot wants K3 judged by whether it can keep working across a long technical task, not just answer a short prompt.

The benchmark signal

Early third-party results make the launch harder to ignore.

Vals AI says Kimi K3 scored 74.70% on the Vals Index, placing second out of 38 models overall in its current table. Vals lists K3's strongest component results as 95.10% on its SWE-bench Verified subset and 91.27% on its Vibe Code Bench subset.

Artificial Analysis gives Kimi K3 a score of 57 on its Intelligence Index, placing it well above the average of comparable models in that dataset. Its write-up frames K3 as comparable to high-end closed models while still behind the very top systems.

Arena's Frontend Code result is the most shareable developer benchmark: Kimi-K3 reached No. 1 in Frontend Code Arena with 1,679 points, ahead of Claude Fable 5 in that leaderboard snapshot.

Benchmarks are not product truth. They are also sensitive to harnesses, model settings, and task selection. Moonshot's own footnotes say different models were evaluated under different agentic harnesses depending on the benchmark. Still, the convergence of official case studies, independent indexes, and frontend coding results gives K3 a serious launch signal.

What is available now

Kimi K3 is already live across Moonshot's own surfaces:

  • Kimi.com and the Kimi mobile apps for general use
  • Kimi Work desktop app, version 3.1.0 or later, for knowledge-work workflows
  • Kimi Code, where users can select K3 through the /model command
  • Kimi API, using the kimi-k3 model name
  • Kimi Enterprise for organizations that need managed privacy and member controls

The important caveat is the open-weight timing. Moonshot says it is working with inference partners and open-source maintainers before the full weights are released by July 27. Until that happens, K3 is available to use, but the open-model ecosystem cannot fully test deployment, fine-tuning, hosting cost, and independent reproducibility.

Why this matters

Kimi K3 sharpens the open-versus-closed AI race in three ways.

First, it raises expectations for open-weight scale. A 2.8-trillion-parameter open model is not easy to run, but the very existence of one changes what developers and enterprises expect from open AI labs.

Second, it pushes the competition toward agentic work. K3 is being sold around long tasks, codebases, visual feedback, spreadsheets, dashboards, research, and software workflows. That is where the money is moving: not better autocomplete, but models that can hold context, use tools, and keep improving a deliverable.

Third, it puts pricing pressure on frontier models. Even if K3 does not beat the top closed systems everywhere, it creates a credible alternative for teams that care about coding, long context, and API cost.

The risk is that early excitement runs ahead of practical reality. Huge open models still need serious serving infrastructure. Enterprise buyers will want security reviews. Developers will want independent evals after the weights land. And global AI politics will keep shaping how companies think about Chinese model providers.

But the launch is still important. Kimi K3 is a clear signal that the frontier is no longer only a closed-model contest between a few U.S. labs. Open-weight models are getting larger, more agentic, and more directly aimed at the workflows where professionals actually spend money.