On July 16, 2026, Moonshot AI shook the artificial intelligence landscape by officially launching Kimi K3. Positioned as the company’s new flagship model, Kimi K3 is officially the world’s largest open-weight AI model to date, boasting an unprecedented 2.8 trillion parameters structured as a Mixture of Experts (MoE).
Kimi K3 represents a significant milestone in open-source AI, offering native multimodality, advanced reasoning capabilities, and an expansive 1 million token context window that rivals closed-source state-of-the-art systems.
Architectural Scaling: 2.8T Mixture of Experts (MoE)
Scaling a model to 2.8 trillion parameters requires intelligent design to remain computationally viable. Moonshot AI achieved this through a highly optimized Mixture of Experts architecture coupled with a new attention mechanism:
graph TD
Input[Input Token] --> Router{MoE Router}
Router -->|Active Weights| Expert1[Expert 1: Coding]
Router -->|Active Weights| Expert2[Expert 2: Reasoning]
Router -->|Inactive Weights| Expert3[Expert 3: Language]
Expert1 & Expert2 & Expert3 --> Output[Aggregated Output]At its core, Kimi K3 uses a hybrid linear attention mechanism named Kimi Delta Attention (KDA) and Attention Residuals to dramatically lower memory bandwidth requirements during long-context inference. While the model has 2.8T total parameters, only a small fraction of these weights are active per query, maximizing throughput and reducing cost.
Frontier Capabilities & API Pricing
Kimi K3 natively supports vision inputs alongside text, allowing it to interpret images, schematics, and screenshots directly. The model is available in two main configurations:
- K3 Max: Tailored for general conversational interface and agentic logic.
- K3 Swarm Max: Designed for massively parallel workflows and swarm orchestrations.
Model Cost Comparison Matrix
| Model | Parameter Access | Input Cost / Million | Output Cost / Million | Context Window |
|---|---|---|---|---|
| Kimi K3 | Open-Weight / API | $3.00 | $15.00 | 1,000,000 |
| GPT-5.6 Soul | Proprietary API | $6.00 | $30.00 | 1,000,000 |
| Claude Fable 5 | Proprietary API | $10.00 | $50.00 | 200,000 |
While Kimi K3 is priced higher than standard developer-focused open-weight models, it undercuts the leading proprietary flagships (GPT-5.6 Soul and Claude Fable 5) by more than 50% on API costs. Moonshot AI intends to publish the full open-weights on July 27, 2026, enabling developers to self-host and customize the model directly.
Image Asset Specifications
- Hero Image:
- Prompt: “Premium clean developer editorial illustration. A glowing abstract neural network cloud with Mixture-of-Experts routers, light background, pastel colors (blue, lavender, white), vector-like clean gradients.”
- Filename: “kimi-k3-news-hero.png”
- Alt text: “Kimi K3 open-weight model announcement illustration”
- Caption: “Kimi K3 scales open-weights to 2.8 trillion parameters using Mixture-of-Experts.”
- Placement: Top of page
- Purpose: Hero image introducing Kimi K3 release
- Aspect ratio: 16:9
- Supporting Visual 1:
- Prompt: “Minimalist UI displaying routing paths and database layers, clean vector card design on a soft grey canvas with pastel violet details.”
- Filename: “kimi-k3-moe-structure.png”
- Alt text: “Mixture of Experts routing schematic”
- Caption: “Kimi K3 uses a Mixture-of-Experts (MoE) routing layout to selectively activate parameters.”
- Placement: Under ‘Architectural Scaling’ section
- Purpose: Clarify MoE parameter routing logic
- Aspect ratio: 4:3
- Supporting Visual 2:
- Prompt: “Modern clean comparison table styling, financial data visualization with soft blue and green graphs, editorial style.”
- Filename: “kimi-k3-pricing-comparison.png”
- Alt text: “API token pricing comparison bar chart”
- Caption: “Kimi K3 provides frontier-class reasoning at a fraction of the cost of proprietary counterparts.”
- Placement: Under ‘Frontier Capabilities & API Pricing’ section
- Purpose: Highlight Kimi K3 cost savings compared to closed-source flagships
- Aspect ratio: 4:3
Related Reading & Workflows
- For long-context engineering guidelines, read our guide on Maximizing Kimi K3: Best Practices for 1M Token Context Windows.
- For a head-to-head performance breakdown against open-weight models, see Kimi K3 vs DeepSeek R1: Architecture, Context, Coding and Deployment Compared.
- For comparison against proprietary models, see Kimi K3 vs Claude Fable 5 vs GPT-5.6.
