Kimi K3: Open Frontier Intelligence

We are thrilled to announce the launch of Kimi K3, our most sophisticated model to date. Designed to push the boundaries of frontier intelligence, Kimi K3 is engineered specifically for high-complexity reasoning, professional knowledge work, and long-horizon software engineering.

🚀 Core Specifications

Kimi K3 is a powerhouse of scale and efficiency, featuring:

• Parameter Count: 2.8 Trillion (The first open 3T-class model).
• Context Window: 1 million tokens.
• Native Capabilities: Integrated vision and advanced reasoning.
• Architectural Foundations: Built upon Kimi Delta Attention (KDA) and Attention Residuals (AttnRes).

While Kimi K3 is the absolute best still trails slightly behind the top-tier proprietary giants—specifically Claude Fable 5 and GPT 5.6 Sol—it has demonstrated frontier-level mastery across our internal benchmarks, consistently eclipsing all other tested open-weight models.

---

🛠 Availability & Deployment

Kimi K3 is accessible starting today via the following platforms:

• Kimi.com
• Kimi Work
• Kimi Code
• Kimi API

Current Status:

• Default: Max thinking effort.
• Upcoming: Low-effort mode.
• Upcoming: High-effort mode.

We are collaborating with open-source maintainers and inference partners to ensure a seamless and stable integration across the wider AI ecosystem. Technical reports covering training and architecture are forthcoming.

---

🧠 The Architecture of a 3T-Class Model

Kimi K3 represents a milestone in scaling. For 9 of the last 12 months, Kimi has defined the ceiling for open-model size. To make a 2.8T parameter model viable, we introduced two critical updates: Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), which optimize information flow across depth and sequence length.

Scaling Efficiency

By utilizing a Stable LatentMoE framework, we scaled Mixture of Experts (MoE) sparsity to activate only 16 out of 896 experts.

The resulting efficiency gain can be expressed as: $\text{Scaling Efficiency}_{K3} \approx 2.5 \times \text{Scaling Efficiency}_{K2}$

Structural Flow

---

💻 Frontier Coding & Engineering

Kimi K3 is not just a chatbot; it is an agentic engineer. It can manage massive codebases, operate terminal tools, and maintain long-term project coherence with minimal human intervention. Its native vision allows it to iterate on CAD, frontend development, and game design by analyzing live screenshots.

Case Study 1: GPU Kernel Optimization

We challenged Kimi K3 to optimize GPU kernels (including KDA and AttnRes) on NVIDIA H200 and alternative GPGPUs.

Case Study 2: Building "MiniTriton"

Kimi K3 autonomously developed MiniTriton, a compact compiler.

• Architecture: Tile-level IR $\rightarrow$ MLIR $\rightarrow$ PTX codegen.
• Performance: Matches or exceeds Triton and torch.compile on specific workloads.
• Validation: Successfully trained nanoGPT with a loss curve nearly identical to the reference.

# Conceptual logic of MiniTriton's pipeline
def compile_kernel(dsl_code):
    ir = transform_to_tile_ir(dsl_code)
    optimized_ir = apply_optimization_passes(ir)
    ptx_code = generate_ptx(optimized_ir)
    return runtime_execute(ptx_code)

---

🎨 Digital Creation & Hardware Design

Interactive Experiences

By combining 3D reasoning and vision, Kimi K3 creates a "vision-in-the-loop" workflow. It writes code, views the rendered screenshot, and refines the output in real-time.

Autonomous Chip Design

In a stunning display of agentic capability, Kimi K3 designed a chip to run a nano-version of its own architecture.

• Timeline: 48-hour autonomous run.
• Tools: Open-source EDA / Nangate 45nm library.
• Specs:
  • Area: $4\text{ mm}^2$
  • Clock: $100\text{ MHz}$
  • Throughput: $> 8,700\text{ tokens/s}$
  • Components: $1.46\text{M cells}$, $0.277\text{ MB SRAM}$, $\text{INT4 MAC array}$.

"A chip built by a model, for a model."

---

🔬 Coding for Scientific Research

Kimi K3 bridges the gap between theoretical papers and executable code. It can synthesize complex research workflows autonomously.

Example: Computational Astrophysics To reproduce the I–Love–Q universal relations, Kimi K3:

1. Analyzed and cross-referenced over 20 scientific papers.
2. Implemented the necessary computational models.
3. Validated the results.

Efficiency Gain: $\text{Human Researcher Time (1-2 weeks)} \rightarrow \text{Kimi K3 Time (2 hours)}$