Moonshot AI, a Beijing-based startup, has released what it describes as the world’s largest open-source AI model, Kimi K3. The company announced Thursday that the model contains 2.8 trillion parameters and a context window of 1 million tokens, positioning it as the first open-source model to reach the 3 trillion parameter range. The claim, if verified, would surpass the scale of existing open-source models like Meta’s Llama 3 (405 billion parameters) and DeepSeek’s V2 (236 billion parameters), though independent verification of the parameter count has not yet been published.

According to Moonshot AI, Kimi K3 is designed for long-context programming, intelligent information retrieval, and reasoning tasks. The company stated that in internal benchmarks, the model achieved state-of-the-art results, outperforming several other tested systems. However, they acknowledged that Kimi K3’s overall performance still trails behind leading proprietary models, such as OpenAI’s GPT-4 and Anthropic’s Claude 3, which remain the top performers in general-purpose evaluations. No specific benchmark scores or comparison methodologies were provided in the announcement.

The release follows a trend of Chinese AI labs pushing the boundaries of model scale in open-source development. Moonshot AI, which has raised significant venture capital, faces the challenge of proving that raw parameter count translates into practical performance gains. Critics note that without transparent benchmarks or peer review, claims of 2.8 trillion parameters in an open-weight model should be treated with caution. The company has not yet released the model weights or a technical paper, leaving the AI community to await further details before validating the achievement.

Source: RIA Novosti