China's Moonshot AI Unveils Kimi K3 - A 2.8 Trillion-Parameter Open-Weight Model That Rivals GPT-5.6 Sol and Claude Fable 5

Moonshot AI has released Kimi K3, a 2.8 trillion-parameter open-weight model that becomes the largest openly available AI system ever built. The Chinese lab claims the model competes with frontier proprietary systems like GPT-5.6 Sol and Claude Fable 5 across coding, knowledge work, and reasoning benchmarks, with full weights to be released by July 27, 2026.

Kimi K3 is not just big. It is the first open model to cross the 3-trillion-parameter threshold and the latest in a sustained push by Moonshot AI that has seen Kimi models set the upper bound of open-model sizes for nine of the past twelve months. The model uses a Mixture of Experts architecture with 896 total experts, activating 16 per forward pass, making it both massive and computationally practical for deployment.

Architecture: Kimi Delta Attention and Attention Residuals

Kimi K3 is built on two novel architectural innovations. Kimi Delta Attention (KDA) provides an efficient foundation for scaling attention across very long sequences, while Attention Residuals (AttnRes) selectively retrieves representations across model depth rather than accumulating them uniformly. Together, they form the backbone of a model designed to scale well beyond the trillion-parameter regime.

Moonshot AI says these structural changes, combined with refined training and data recipes, yield an approximate 2.5x improvement in overall scaling efficiency compared to Kimi K2. The model also uses Stable LatentMoE with Quantile Balancing to eliminate heuristic routing updates, and Per-Head Muon optimization for adaptive learning at scale.

The model features a 1-million-token context window and native vision capabilities, supporting text, image, and video understanding within a single architecture. This multimodal design lets Kimi K3 reason across screenshots, diagrams, and video frames without separate vision encoders.

Benchmarks: Competitive with Frontier Proprietary Models

Kimi K3 trails Claude Fable 5 and GPT-5.6 Sol on aggregate benchmarks, but Moonshot AI reports it consistently outperforms every other tested model including Claude Opus 4.8, GPT-5.5, and GLM-5.2 across coding, knowledge work, and reasoning evaluations.

In coding, Kimi K3 achieved competitive results on GPU kernel optimization tasks, matching Fable 5 on benchmarked workloads and substantially outperforming Opus 4.8 and GPT-5.6 Sol. In a striking demonstration of autonomous capability, an early version of Kimi K3 handled the majority of the team's own kernel optimization work during late-stage development.

The model also built MiniTriton, a compact Triton-like compiler with its own tile-level IR layer, optimization passes, and PTX code-generation pipeline. Across supported roofline benchmarks, MiniTriton delivers performance on par with or better than Triton and torch.compile, validating that Kimi K3 can build coherent end-to-end engineering systems rather than isolated code snippets.

Open Weight Release and Pricing

Full model weights will be released by July 27, 2026, alongside a forthcoming technical report detailing architecture, training methodology, and evaluations. Moonshot AI is working with inference partners and open-source maintainers to align technical details and ensure reliable rollout across the ecosystem.

For those accessing the model via API, pricing is set at $0.30 per million tokens for cache-hit input, $3.00 per million tokens for cache-miss input, and $15.00 per million tokens for output. Powered by Mooncake's disaggregated inference architecture, the official Kimi API achieves a cache hit rate above 90 percent in coding workloads, making the effective cost significantly lower than the headline output price suggests.

Moonshot AI recommends deploying Kimi K3 on supernode configurations with 64 or more accelerators for optimal inference performance. The model uses quantization-aware training from the SFT stage onward, applying MXFP4 weights with MXFP8 activations for broad hardware compatibility.

What This Means for the AI Landscape

Kimi K3 changes the competitive dynamics of the open-weight model space dramatically. Prior to this release, the largest open models sat in the hundreds-of-billions range. A 2.8 trillion-parameter open model means that organizations building on AI can now access frontier-level capability without being locked into a proprietary API. The open-source ecosystem just gained a serious contender at the top of the benchmark charts.

For founders and builders, the implications are significant. Open-weight models at this scale reduce dependency on single-vendor API pricing and availability. If Moonshot AI follows through on its weight release promise, the cost structure for running large-scale inference could shift meaningfully, especially in markets where data sovereignty or latency demand on-premise deployment.

The broader signal, however, is about China's position in the AI race. Kimi K3 demonstrates that Chinese AI labs are producing truly frontier-level research, not just catching up. With open-weight distribution, these capabilities are now globally accessible, accelerating the pace of AI development regardless of geopolitical boundaries.

Kimi K3 is available now via the Kimi app (iOS, Android, HarmonyOS), the Kimi desktop app (Windows and Apple Silicon Macs), and the official API. The weight release on July 27 will be one of the most consequential open-source AI events of the year.