Kimi K3 2.8T-A50B: Largest Open Model Ever Released, Opus 4.8-Class at Sonnet 5 Pricing

On July 16, 2026, Moonshot AI officially released Kimi K3, with a parameter scale of 2.8 trillion, making it the largest open-source model ever. This number far exceeds previous open-source models (e.g., Llama 3 405B) and even surpasses some closed-source models (e.g., GPT-4's 1.8T parameters). The release of Kimi K3 marks a new magnitude in the scale of open-source large models.

Before the release, Moonshot AI released an artistic teaser video hinting at its technical breakthrough. Kimi K3 performs excellently on multiple benchmarks, especially in long-context understanding and complex reasoning tasks. The company claims its performance is 'only half a step away from Fable 5 and GPT-5.6 Sol.' This phrasing suggests that Kimi K3 is close to the current closed-source top level, but there is still a slight gap.

Moonshot AI compares Kimi K3 with GPT-5.6 Sol and Claude Fable 5, emphasizing its 'half-step' gap. Specifically, Kimi K3 scores close to them on standard benchmarks like MMLU and HumanEval, but may be slightly inferior in some complex reasoning or safety tests. The company has not released complete comparison data but hints that the gap mainly lies in 'extreme long-tail tasks' and 'alignment optimization.'

Notably, the open-source nature of Kimi K3 means the community can reproduce, fine-tune, and even improve the model, which may narrow the gap in the short term. However, closed-source models typically have more refined RLHF and larger inference infrastructure, so the 'half-step' gap may persist. Additionally, the training cost for a 2.8T-parameter model is estimated to be hundreds of millions of dollars, making it difficult for ordinary developers to replicate.

The open-source release of Kimi K3 is a major boon for the AI community. Developers can use it for domain fine-tuning, research on model interpretability, or even building specialized applications. Moonshot AI may hope to attract community contributions to accelerate model iteration, similar to Meta's Llama strategy.

However, the actual deployment of a 2.8T-parameter model requires massive GPU resources (e.g., thousands of H100s), limiting its use by small and medium-sized teams. Moonshot AI may provide an API or quantized versions to lower the barrier, but specific plans have not been announced. Additionally, open-source models may face risks of misuse, such as generating harmful content, requiring Moonshot AI to provide corresponding safety mechanisms.

The release of Kimi K3 further blurs the line between open-source and closed-source models. Previously, open-source models lagged behind closed-source ones by 1-2 years in scale, but Kimi K3 almost catches up to the level of GPT-5.6 Sol (released in 2026). This suggests that the open-source community is catching up quickly, but closed-source models (e.g., GPT-5, Gemini Ultra 2) may have already entered a higher dimension.

Moonshot AI's choice to open-source rather than keep it closed may be driven by brand building, community ecosystem, or policy considerations. In the Chinese market, open-source models help establish technical influence and attract developers to its platform. However, in the long run, commercial closed-source models can maintain their lead through continuous investment, while open-source models rely on community contributions and financial support.