Moonshot AI dropped Kimi K3 on July 16 without a launch event or press conference. The Chinese AI startup simply updated kimi.com, and anyone could start using it immediately. Kimi K3 is a 2.8 trillion parameter Mixture-of-Experts model, making it the largest open-weight AI model released to date.
What Makes Kimi K3 Different
The model uses a Mixture-of-Experts architecture, meaning it does not activate all 2.8 trillion parameters for every query. Instead, it routes inputs to the most relevant subsets of parameters, which keeps inference costs manageable while maintaining the breadth of a massive model.
Key specifications:
- 2.8 trillion total parameters (open weights)
- 1 million token context window
- Native vision capabilities (multimodal from the ground up)
- Kimi Delta Attention (KDA) for efficient long-context processing
- Open-source release on Hugging Face
Kimi Delta Attention Explained
Moonshot developed a custom attention mechanism called Kimi Delta Attention (KDA) to handle the 1 million token context window efficiently. Standard transformer attention scales quadratically with context length, which makes million-token contexts impractical with vanilla architectures. KDA uses delta-based compression to reduce the memory and compute requirements while preserving the model ability to retrieve information from distant parts of the context.
Performance Benchmarks
Moonshot claims Kimi K3 outperforms several closed-source competitors on reasoning, coding, and long-context benchmarks. The model reportedly uses 30% fewer reasoning tokens than comparable models on coding tasks, which translates directly to lower inference costs for developers deploying it at scale.
On mathematical reasoning benchmarks, Kimi K3 scored competitively with GPT-5.6 Sol and Claude Opus, though independent verification is still pending as researchers work through the full evaluation suite.
The Open Weight Advantage
By releasing the weights openly, Moonshot is betting that adoption matters more than keeping the model proprietary. Companies can fine-tune Kimi K3 for domain-specific tasks, run it on their own infrastructure for data privacy, and avoid API lock-in.
This positions Kimi K3 as a direct competitor to Meta Llama models, which have dominated the open-weight space. The 2.8 trillion parameter count dwarfs Llama 3.1 405B, though the MoE architecture means actual compute requirements are not directly comparable.
Moonshot AI Background
Moonshot AI was founded in 2023 and has raised over 1 billion dollars in funding. The company previously gained attention with its Kimi chatbot, which became popular in China for its long-context capabilities. Kimi K3 represents a significant leap in ambition, positioning the startup as a serious contender in the global AI race alongside OpenAI, Anthropic, Google, and Meta.
Frequently Asked Questions
What is the Kimi K3 model?
Kimi K3 is a 2.8 trillion parameter Mixture-of-Experts language model released by Moonshot AI on July 16, 2026. It features a 1 million token context window, native vision support, and is fully open-weight.
How does Kimi K3 compare to GPT-5.6?
Moonshot claims Kimi K3 is competitive with GPT-5.6 Sol on reasoning and coding benchmarks, with 30% fewer reasoning tokens on code tasks. Independent benchmarks are still being compiled.
Where can I download Kimi K3?
Kimi K3 weights are available on Hugging Face. You can also use it directly through the Kimi web interface at kimi.com without any API setup.
Is Kimi K3 free to use?
The weights are open-source, so self-hosting is free (minus compute costs). The kimi.com web interface also offers free access with usage limits, plus a paid tier for higher throughput.
