The Dark Side of the Moon officially open-sources “Kimi K3”! With powerful programming skills, it reached #1 on Hugging Face within half an hour

China’s AI startup Moonshot AI once again shook the global open-source community! Hugging Face CEO Clement Delangue announced on X that Moonshot AI’s latest open-source 2.8 trillion-parameter multimodal model “Kimi K3” has officially launched. In just 30 minutes, it amassed more than 4,000 likes, jumped straight to #1 on the trending list, and set the fastest growth record in the platform’s history.
(Background: US and UK official reports: China’s Kimi K3 security offensive-and-defense capabilities are “clearly behind” the top US AI models)
(Additional background: Is self-hosting AI models cheaper? Cline tested Kimi K3 and found a cost advantage, but netizens argued it’s more expensive than Fable 5 when priced per task)

Table of contents

Toggle

  • The first open-source 3T-class model worldwide, with a million-level context
  • Focused on long-horizon coding and reasoning; benchmark performance rivals frontier models
  • Flexible deployment with a dedicated Kimi K3 commercial license

As competition in artificial intelligence (AI) computing power and model performance heats up, the open-source community welcomed a heavyweight new force. On July 27, Beijing Time, Clement Delangue, co-founder and CEO of Hugging Face, the world’s largest open-source AI community platform, posted on X (formerly Twitter) to share good news, announcing that Moonshot AI’s new model “Kimi K3” has officially arrived on the platform.

What’s astonishing is that after only 30 minutes of launch, Kimi K3 quickly attracted over 4,000 likes, taking the #1 spot on Hugging Face’s “Trending” leaderboard. Delangue marveled that this is the fastest growth the platform has ever seen after a release.

Kimi K3 has been released on HF and already top 1 trending with 4,000+ likes in 30 mins. Fastest release growth ever so far!! pic.twitter.com/2lnbbgXwGD

— clem 🤗 (@ClementDelangue) July 27, 2026

The first open-source 3T-class model worldwide, with a million-level context

According to information on Hugging Face’s official page, Kimi K3 is Moonshot AI’s strongest open-weight native multimodal agent model to date. Its total parameter count reaches 2.8 trillion (2.8T), and industry insiders call it the first open-source model to reach the 3T-class level worldwide.

In terms of technical architecture, Kimi K3 uses a new Kimi Delta Attention (KDA) and Attention Residuals architecture, and greatly improves scaling efficiency through the Stable LatentMoE framework. Out of 896 expert networks, only 16 are activated each time, keeping the effective activated parameter count at 104B. Compared with its predecessor Kimi K2, its overall scaling efficiency is improved by about 2.5 times. In addition, it natively supports an ultra-large context window of up to 1 million tokens, allowing it to easily handle text and image inputs at the same time.

Specialized for long-horizon coding and reasoning; benchmarks rival frontier models

Designed specifically for cutting-edge intelligence, Kimi K3 is especially strong in knowledge work, complex reasoning, and long-horizon coding. The official technical blog states that with minimal human supervision, the model can autonomously browse large codebases and perform tasks such as GPU kernel optimization, compiler development, and even difficult chip design.

In various demanding benchmark tests, Kimi K3 shines when using the maximum thinking mode. It not only scored an excellent 93.5 on GPQA Diamond and 88.3 on Terminal-Bench 2.1, but its combined agent-style tasks and visual capabilities already put it on par with top closed-source models like Claude Fable 5 and GPT-5.6 Sol.

Flexible deployment with a dedicated Kimi K3 commercial license

To improve hardware compatibility and runtime efficiency, Kimi K3 introduced native MXFP4 quantization-aware training starting from the SFT (supervised fine-tuning) stage. It can also seamlessly support mainstream open-source inference frameworks such as vLLM and SGLang, so developers can deploy immediately after downloading the weights.

Currently, the model’s full source code and weights have been released via a custom “Kimi K3 License,” allowing developers to conduct research, fine-tuning, and further development. However, the official also explicitly states that if a model API service provider wants to use it commercially with annual revenue exceeding $20 million, it must sign a separate commercial agreement with Moonshot AI. With Kimi K3 going viral, the global AI industry is expected to spark a new wave of innovative applications built on this model.

View Original
This page may contain third-party content, which is provided for information purposes only (not representations/warranties) and should not be considered as an endorsement of its views by Gate, nor as financial or professional advice. See Disclaimer for details.
  • Reward
  • Comment
  • Repost
  • Share
Comment
Add a comment
Add a comment
No comments
  • Pinned