Credit: Shutterstock
Kimi K3’s full weights are now public, making it the largest open-weight AI model released so far.
Moonshot AI published the complete model weights for Kimi K3 on Hugging Face today (00:00 UTC July 27). This gives developers and researchers direct access to a 2.8 trillion-parameter model for the first time.
The release closes out an 11-day rollout that began on July 16, when the Beijing-based startup first introduced K3 through its API and consumer app at the World AI Conference in Shanghai.
Until today, K3 was only available as a hosted model. Anyone who wanted to use it had to go through Moonshot’s API or its kimi.com web app. Now that the full weights are downloadable, organizations with enough hardware capability can run the model on their own servers, study how it works, and fine-tune it for specific tasks.
Moonshot built K3 on a new architecture called Kimi Delta Attention, combined with a technique the company refers to as Attention Residuals. Moonshot says this combination allows the model to handle long coding sessions and complex reasoning tasks more efficiently than earlier Kimi models. K3 also includes built-in agentic features, such as tool calling, web browsing, and multi-step task planning.
According to Moonshot’s own benchmark results, K3 trails Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 Sol on overall performance. However, the company says it outperformed GPT-5.5 and Claude Opus 4.8 on certain coding and agent benchmarks. Since these results come from Moonshot itself, the Open-Weight Release should make it easier for independent researchers to test the claims directly.
Running K3 is not easy.
The model comes in at roughly 594 gigabytes in its released format, and Moonshot estimates that at least 8 high-end GPUs, at the level of an H100 or better, are needed to load it. This means self-hosting is not realistic for most individual developers, though it remains possible for companies and research labs with significant computing budgets.
The license is expected to follow the Modified MIT terms Moonshot used for its earlier K2 models. The company has not yet published the final license file, so anyone planning to build a commercial product on K3 should confirm the exact terms once the weights and documentation are fully live.
Independent Testing Shows Mixed Results
In the days before the open-weight release, several independent groups tested K3 through its hosted API. Their findings differ depending on what was measured.
- The UK AI Security Institute and the US Center for AI Standards and Innovation ran a joint safety evaluation. They found that K3 performs below frontier US models on offensive cyber tasks. In a test involving a simulated 32-step network attack, K3 completed an average of 17 steps, compared to about 28.5 for the most capable US models tested. The same evaluation noted that K3’s safety guardrails did not stop it from attempting cyber exploit development when instructed to try.
- Belgian security firm Aikido tested K3 against 26 recently discovered software vulnerabilities. The model correctly identified 23 of them, a detection rate close to OpenAI’s mid-tier GPT-5.6 Terra, at roughly a quarter of the cost. Aikido described K3 as the strongest open-weight model currently available for this type of cybersecurity work.
What About the US-China AI Story?
K3’s rise has revived an ongoing debate. US officials and Anthropic have accused Moonshot of training K3 using outputs from American AI models, a practice known as Distillation. Moonshot has denied this.
The model has also raised new questions about whether building frontier AI truly requires the scale of computing power and advanced chips that US export controls have tried to limit. Two days after its initial launch, demand for K3 was high enough that Moonshot temporarily paused new subscriptions to protect service for existing users.
With the weights now public, independent developers and researchers can attempt to reproduce Moonshot’s benchmark claims, identify where the model performs well or falls short, and build their own tools on top of its architecture. How the developer community responds to K3 in the coming weeks will provide a clearer picture of its actual capabilities than Moonshot’s own marketing has offered so far.
