The world's largest open-source model by parameters, even Elon Musk gave it a thumbs up! Moonshot AI's Kimi K3 open-sourced with 2.8 trillion parameters.

Recently, Moonshot AI officially released the Kimi K3 model weights and technical report, and simultaneously open-sourced three key infrastructure technologies that support the training of this model: MoonEP, FlashKDA, and AgentEnv.

According to the introduction, the three technologies played an important role in the training process of Kimi K3, providing foundational support for the efficiency, stability, and agent task execution of large-scale model training.

It is understood that Kimi K3 is currently the most capable model from Moonshot AI. This model adopts a Mixture of Experts (MoE) architecture, with a total parameter scale of 2.8 trillion, possesses native visual understanding capabilities, and supports a context window of up to 1 million tokens.

Among them, AgentEnv is an agent sandbox system jointly developed by Moonshot AI and KVCache.ai, mainly used for running Agent environments at scale.

This system provides a high-fidelity, strongly isolated running environment for Kimi K3 post-training, and supports functions such as fast snapshot, restore, and fork, meeting the needs of large-scale parallel agent workflows and training tasks.

Kimi K3 was officially released on July 17, mainly targeting scenarios such as long-range programming, knowledge work, and complex reasoning.

This model is currently the open-source model with the largest parameter scale, and its release quickly attracted attention from the AI industry both domestically and internationally.

Tesla CEO Elon Musk also commented on the relevant evaluation report, calling it "impressive."

Within hours after its release, Kimi K3 topped the AI code tool evaluation leaderboard Arena, becoming the first Chinese large model to achieve first place on this leaderboard.

Angelopoulos, the operations lead of the leaderboard, stated that the release of Kimi K3 may force investors to re-evaluate the entire AI industry and trigger a reshuffle of the capital market.

This release of model weights, technical report, and training infrastructure means that Moonshot AI is no longer just open-sourcing the model itself, but is further showcasing the large model training and agent post-training system to developers and research institutions.

The continued open-sourcing of related technologies is expected to lower the threshold for ultra-large-scale model research and agent application development, and drive the domestic open-source model ecosystem to extend from model capability competition to competition in underlying engineering capabilities and toolchains.