Moonshot AI drops Kimi K3, a 2.8T open-weights giant

Moonshot AI has released Kimi K3, a 2.8-trillion parameter open model that marks the industry's first open 3T-class frontier system. Designed natively with vision and a 1-million-token context window, it targets sustained, long-horizon coding sessions, multi-step knowledge workflows, and autonomous technical research.

What it can do

K3 focuses heavily on long-duration agentic work with minimal human steering. It can navigate large code repos, operate terminal environments, and use visual feedback from live screenshots to debug game development, frontends, and CAD projects.

Its technical capabilities shine in complex systems engineering. In internal evaluations, K3 spent 24 hours in an isolated sandbox profiling and optimizing low-level GPU kernels, outperforming GPT 5.6 Sol, GPT 5.5, and Opus 4.8 while staying competitive with Claude Fable 5. Moonshot notes that early versions of K3 handled most of the team's internal kernel optimization tasks near the end of training.

It can also build infrastructure from scratch. K3 created MiniTriton, an end-to-end GPU compiler featuring its own intermediate representation and PTX codegen pipeline, capable of driving full nanoGPT training to stable convergence. In another 48-hour hardware sprint, it autonomously designed and validated an ASIC chip using open-source EDA tools, achieving over 8,700 tokens per second in simulation.

On external evaluations, K3 scored a 57 on the AA Intelligence Index, landing just behind proprietary frontrunners Fable 5 at 60 and Sol at 59.

Architecture and hardware requirements

K3 is powered by Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), which streamline attention scaling across long contexts and deep network layers. It runs on a Stable LatentMoE structure featuring 896 experts while activating 16 per forward pass. Along with Quantile Balancing routing and Per-Head Muon optimization, Moonshot claims a 2.5x jump in scaling efficiency over Kimi K2.

Running a model of this magnitude locally is demanding. K3 uses quantization-aware training with MXFP4 weights and MXFP8 activations. Because expert-parallel routing relies heavily on communication bandwidth, Moonshot recommends deploying the model across supernodes packed with at least 64 hardware accelerators.

Full model weights are scheduled for public release on July 27, 2026, alongside technical documentation and a dedicated prefix caching pull request for vLLM.

You do not need a supercomputer to test it today. K3 is live on kimi.com, the Kimi Work 3.1.0 desktop client, the Kimi Code CLI, and through Moonshot's official API. Defaulting to maximum thinking effort, API pricing sits at $0.30 per million tokens for cached input, $3.00 for uncached input, and $15.00 for output, matching Claude 5 Sonnet rates.

Why it matters

Kimi K3 substantially closes the performance gap between open-weight technology and closed-source flagships. While industry consensus has long assumed open source sits roughly 6 to 12 months behind labs like OpenAI and Anthropic, K3 shows that gap may now be closer to a single release iteration in areas like heavy systems coding and math-heavy reasoning.

For everyday developers, self-hosting a 2.8T architecture remains impractical due to cluster hardware requirements, making the API the rational testbed today. But for organizations with serious on-premise compute, the July 27 weights offer something rare: near-frontier intelligence on an open foundation without the privacy trade-offs of proprietary cloud APIs.