Kimi K3: The 2.8T Open-Weight Model Reshaping the AI Landscape

On July 16, 2026, Moonshot AI released Kimi K3 — the world's first open-weight model with over 2.8 trillion parameters. Built on a novel architecture combining Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), K3 is not just another large language model. It is a native multimodal agentic model designed for long-horizon coding, deep knowledge work, and complex reasoning — and it is available for anyone to download, inspect, and deploy.

What Makes Kimi K3 Different

Kimi K3 is a Mixture-of-Experts (MoE) model with 2.8 trillion total parameters, but only 104 billion active per inference step. It activates 16 out of 896 experts per token, achieving an approximate 2.5x improvement in scaling efficiency over its predecessor, Kimi K2. This sparse activation means K3 delivers frontier-level intelligence while remaining practical to deploy — a balance that has historically been elusive in open-weight models.

Intelligence Leaderboard

According to Artificial Analysis, Kimi K3 scores 57 on the Intelligence Index, placing it well above the median of 25 for open-weight models of similar size. It ranks as the #1 open-weight model overall, surpassing GLM-5.2 (51) and MiniMax-M3 (44). Among all models — including proprietary ones — it sits in the top tier alongside Claude Opus 5, Claude Fable 5, and GPT-5.6 Sol.

Across the board, Kimi K3 competes with — and in many cases surpasses — leading proprietary models. Its performance on BrowseComp (91.2%) and Terminal-Bench 2.1 (88.3%) is particularly striking, demonstrating strong agentic and coding capabilities.

Agentic by Design

Kimi K3 was built from the ground up as an agentic model. Unlike many models that treat tool use and reasoning as post-training additions, K3 was trained in preserved thinking history mode — meaning it natively understands how to reason, call tools, and iterate without losing context.

This approach pays off in benchmarks like GDPval-AA v2 (agentic real-world tasks), where K3 scores 1686 Elo, and AA-Briefcase (agentic knowledge work), where it achieves 1548 Elo. It also excels at MCP (Model Context Protocol) tasks with a verified score of 94.5%, and achieves 76.5% on Toolathlon-Verified — a measure of multi-tool orchestration.

For developers, this means Kimi K3 can sustain long engineering sessions with minimal human oversight. It navigates massive code repositories, orchestrates terminal tools, and can handle tasks ranging from GPU kernel optimization to compiler development to game development with vision feedback.

Open Weights, Many Suppliers

One of the most significant aspects of Kimi K3 is its openness. The full model weights are available on Hugging Face under the Kimi K3 License, making frontier intelligence accessible for research, deployment, and further innovation. Because it is open-weight, K3 is also available through9 different API providers, each offering different performance and pricing.

This competitive landscape means users can choose based on their priorities: Makora and Fireworks offer the fastest inference (170 and 163 tokens/second), while Parasail and Makora provide the lowest pricing at $2.00/1M input tokens. The first-party Kimi API offers a solid baseline at $3.00/$15.00 per 1M tokens with full feature support.

Pricing and Value

At $3.00 per 1M input tokens and $15.00 per 1M output tokens from the first-party API, Kimi K3 is positioned as a premium open-weight model. For comparison, this is roughly 3x the cost of Kimi K2.5 ($0.35/$1.89) but delivers significantly higher intelligence — a 57 vs 25 Intelligence Index score. For cost-sensitive workloads, providers like Parasail and Makora offer the same model at $2.00/$10.00, a 33% discount.

Kimi K3 also supports prompt caching with a 90% discount on cached tokens ($0.30/1M), making it more economical for repeated query patterns common in agentic workflows.

How to Get Started with Kimi K3

Because Kimi K3 uses an OpenAI-compatible API, getting started is straightforward. You can call it through any OpenAI client library by pointing it at your preferred provider's endpoint. The model supports vision input, structured output (JSON mode), function calling, and context caching.

import openai

client = openai.OpenAI(
    base_url="https://api.krxlab.com/v1",  
    api_key="sk-krx-..."
)

response = client.chat.completions.create(
    model="kimi-k3",
    messages=[
        {"role": "user", "content": "Write a Python script to analyze a CSV file"}
    ],
    reasoning_effort="max"
)

print(response.choices[0].message.content)

For agentic use cases, Moonshot recommends using the Kimi Code CLI as the agent framework, which provides native support for long-horizon coding sessions, multi-tool orchestration, and vision-in-the-loop development.

The Bottom Line

Kimi K3 represents a significant milestone in open-weight AI. It is the first model to break the 2.8T parameter barrier while remaining accessible to the community, and its agentic design sets a new standard for what open models can achieve. Whether you are building coding agents, research tools, or enterprise applications, K3 offers frontier-level intelligence with the flexibility of open weights and the choice of 9 different API providers.

The era of open frontier intelligence has arrived — and it runs on Kimi K3.