Moonshot releases Kimi K3, largest open-weight model to date

An open-weight release at genuine frontier scale is the thing worth paying attention to here. According to VentureBeat, Moonshot AI, the Beijing startup backed by Alibaba, released Kimi K3 on July 16, a 2.8-trillion-parameter mixture-of-experts model the company says is the largest open-source AI system ever built, with benchmark scores that put it in the same neighborhood as the top proprietary systems from Anthropic and OpenAI.

The mechanics matter more than the raw parameter count. Only 16 of the model's 896 experts activate per token, roughly 1.8 percent of the pool, so inference costs sit closer to a much smaller dense model. Moonshot pairs that with a 1-million-token context window, native vision, an always-on reasoning mode the company calls 'thinking mode,' and two architectural additions it names Kimi Delta Attention and Attention Residuals. The API is already live, with the full weights scheduled to publish on July 27.

The reported numbers put K3 in close range of the closed frontier. On the GDPval-AA v2 benchmark it scored 1,687, third behind Claude Fable 5 Max at 1,815 and GPT-5.6 Sol Max at 1,747.8, and ahead of Claude Opus 4.8 at 1,600, per coverage from Tom's Hardware. K3 also topped Frontend Code Arena, a 17-place jump from its predecessor K2.6's #18. For anyone building on paid US APIs, a downloadable model playing in that league changes what the fallback plan looks like.

The honest caveats are the ones you would guess. Benchmark leaderboards flatter models trained with them in mind, and what the reporting so far does not give you is the training compute story, the exact license terms of the July 27 weight drop, or independent verification of these scores under real workloads. Take the specifics as reported, not settled. What is settled is the direction and the timing, landing right before the 2026 World Artificial Intelligence Conference in Shanghai. If K3 holds up under outside testing, the leverage US frontier labs get from being the only place to run frontier-class models loosens, and sovereign-cloud and cost-sensitive teams get an option they did not have a week ago.