Moonshot AI Launches Kimi K3 as World's Largest Open AI Model

Chinese startup Moonshot AI has launched Kimi K3, a 2.8-trillion-parameter open-weight Mixture-of-Experts (MoE) model. Positioned as the world’s largest open AI model, Kimi K3 introduces Kimi Delta Attention to deliver up to 6.3x faster decoding in million-token contexts. Built for long-horizon coding, native visual understanding, and always-on reasoning, the model performs competitively with closed frontier systems like Anthropic’s Claude Fable 5, with full model weights scheduled for public release on July 27, 2026.

Inside the 2.8-Trillion-Parameter Stable LatentMoE Architecture

At 2.8 trillion parameters, Kimi K3 is the largest open-weight model available, surpassing previous open-weights models such as DeepSeek V4 Pro (1.6T) and GLM-5 (753B). To make a model of this scale computationally and economically viable, Moonshot AI utilizes a sparse Mixture-of-Experts (MoE) architecture powered by a Stable LatentMoE framework.

Under this setup, the neural network’s processing capacity is divided into specialized sub-units:

• Total Experts: 896 total routing experts.
• Active Experts: 16 active experts per token.
• Active Parameters: Approximately 50 billion parameter equivalent per forward pass.
• Precision Format: Mixed-precision training utilizing MXFP4 weights and MXFP8 activations.

This selective activation keeps computational demand during inference manageable while retaining the massive factual capacity of a 2.8-trillion-parameter system. Developers can run Kimi K3 on self-hosted multi-node GPU clusters, avoiding recurring per-token API fees and keeping proprietary data within local security boundaries. Moonshot AI has announced the full release of Kimi K3’s model weights on July 27, 2026.

Accelerating Long-Context Performance with Kimi Delta Attention and Attention Residuals

The context window of Kimi K3 spans 1 million tokens, allowing the model to hold entire codebases or massive document sets in its active memory. In standard Transformer architectures, processing such long inputs causes a computational bottleneck because the mathematical operations required to calculate token relationships grow quadratically.

To overcome this, Kimi K3 implements two architectural optimizations:

1. Kimi Delta Attention (KDA): A hybrid linear attention mechanism that replaces standard quadratic attention in a subset of layers, retaining expressiveness while dramatically reducing compute across the 1M-token context window.
2. Attention Residuals (AttnRes): A drop-in replacement for standard residual connections that allows layers to selectively retrieve representations from arbitrary earlier layers rather than uniformly accumulating outputs.

According to Moonshot AI, these design improvements result in up to a 6.3x decoding speedup in million-token contexts, ensuring fast generation times even when processing extensive data repositories.

Native Visual Understanding and Always-On Reasoning

Rather than relying on separate vision adapters or pipeline stages, Kimi K3 uses native visual understanding. Text and image inputs are processed simultaneously by a single unified network, allowing the system to preserve spatial relationships and layouts when analyzing diagrams, UI mockups, and charts.

In addition to visual tasks, Kimi K3 operates with an always-on reasoning capability. When prompted, the model generates internal chains of thought to plan its response, evaluate alternative solutions, and self-correct logical errors before returning a final output. By default, Kimi K3 utilizes its maximum reasoning effort, which improves performance on complex mathematical and logical queries.

Developers can call the Kimi K3 API with custom reasoning configurations. Below is an example of an API call requesting a deep reasoning response:

const response = await env.AI.run(
  'moonshotai/kimi-k3',
  {
    messages: [{ content: 'Prove that the square root of 2 is irrational.', role: 'user' }],
    reasoning_effort: 'max',
  },
)
console.log(response)

Long-Horizon Coding Workflows and Autonomous Engineering

Kimi K3 is optimized for agentic operations, which are multi-step, goal-oriented tasks that require planning and self-correction over extended periods with minimal human intervention. When deployed in a software engineering pipeline, the model executes long-horizon workflows through the following steps:

1. Repository Analysis: Navigates a complex codebase to locate relevant modules and map dependencies.
2. Test Generation: Automatically drafts unit and integration test suites for new features.
3. Execution and Monitoring: Runs the test suites within a simulated environment or sandbox.
4. Iterative Debugging: Parses compiler errors or test logs to rewrite problematic segments, repeating the loop until all tests pass successfully.

As a proof of concept, Moonshot AI reported that Kimi K3 successfully designed, optimized, and verified a microchip designed to run a nano-model on its own architecture during an autonomous 48-hour run. Using open-source electronic design automation (EDA) tools on a 45nm library, the model generated a 4 mm² layout containing 1.46 million standard cells and 0.277 MB of SRAM, closing timing at 100 MHz.

Benchmarking Kimi K3: Artificial Analysis and Arena Performance Metrics

Independent evaluation by testing lab Artificial Analysis corroborated Kimi K3’s standing near the proprietary frontier. The model scored 57 on the Artificial Analysis Intelligence Index v4.1, placing it fourth in its comparison class.

On agentic and long-horizon tasks, Kimi K3 achieved the following results in independent benchmarks:

• AutomationBench-AA: Scored 53%, taking the #1 position for agentic SaaS workflows.
• GDPval-AA v2: Reached an Elo rating of 1668, outperforming GLM-5.2 (1514), GPT-5.5 (1494), and Claude Opus 4.8 (1600), while trailing Anthropic’s Claude Fable 5, which scored 1760.
• AA-Briefcase: Achieved an overall Elo of 1547 in long-horizon knowledge work, registering a 732-point improvement over the previous-generation Kimi K2.6 and finishing behind only Claude Fable 5.
• Arena.ai Frontend Code Arena: Debuted at #1, leading all tested proprietary systems, including Claude Fable 5.

While the model shows high intelligence, independent testing also noted that Kimi K3 is highly verbose, consuming 21% fewer output tokens than Kimi K2.6 on index tasks but generating a total of 130 million output tokens across the full evaluation suite—about double the comparison class median of 63 million tokens.

Deployment and API Pricing

For organizations utilizing cloud-hosted deployments, Kimi K3 is available on Kimi.com, Kimi Code, and through API platforms. On OpenRouter, Kimi K3 is priced at $3.00 per million input tokens and $15.00 per million output tokens.

Moonshot’s proprietary hosting infrastructure utilizes a disaggregated system called Mooncake, which separates prefill and decode operations across dedicated node pools. This architecture allows the company to offer cached input pricing at $0.30 per million tokens. Running the full 2.8-trillion-parameter weights locally after the July 27 release will require an aggregate GPU memory of approximately 1.4 TB due to MXFP4 weight quantization.