Moonshot AI, the Beijing-based startup behind the Kimi family of assistants, has released Kimi K3, a 2.8-trillion-parameter model that its makers and independent trackers describe as the largest openly released AI system to date. The company published the full model weights on July 26, one day ahead of a July 27 target it had communicated to developers, and made them available under a permissive, Modified-MIT-style licence that lets companies download, run, and modify the model on their own hardware.
The headline number is deliberately eye-catching, but the way the model is built matters more than its raw size. Kimi K3 is a Mixture-of-Experts system: of its roughly 896 specialised sub-networks, only sixteen are activated for any given token, meaning that just under two percent of the network does the work at any moment. That design keeps the cost of running the model far below what a dense 2.8-trillion-parameter system would demand, and it is the same architectural trick that has let a handful of Chinese labs punch above the compute weight that export restrictions allow them.
Alongside the scale, Moonshot is drawing attention to two homegrown architectural ideas. The first, which it calls Kimi Delta Attention, is a hybrid linear-attention mechanism intended to handle very long inputs more cheaply. The second, Attention Residuals, is described as a drop-in replacement for the residual connections that have anchored transformer models for years, delivering steadier gains as the model scales. The model ships with a one-million-token context window, native visual understanding, and an always-on reasoning mode the company labels "thinking mode."
Why It Matters
For most of the past three years the frontier of AI capability has belonged to a small club of American labs that keep their best models locked behind an API. Kimi K3 is a direct challenge to that arrangement. According to figures compiled by the independent evaluation service Artificial Analysis, the model posted an overall Elo rating of roughly 1,547, trailing only Anthropic's Claude Fable 5 and sitting ahead of OpenAI's GPT-5.5. In a blind frontend-coding arena it ranked first at 1,679 points, edging out Fable 5 in developer testing, and on a broader real-world benchmark spanning dozens of occupations it placed third behind two of the strongest closed American systems.
What makes those numbers consequential is not that a Chinese model beat a Western one on a leaderboard, but that a model anyone can download now sits within striking distance of systems you can only rent. As Forbes observed, K3 is the clearest sign yet of a convergence between open-weight and closed-weight quality, a gap that was measured in years not long ago and is now measured in weeks.
There is also a hardware subtext that industry watchers were quick to note. Moonshot trained and shipped a model of this scale despite the U.S. export controls that limit Chinese access to the most advanced accelerators. As Tom's Hardware framed it, the release is evidence that sparse architectures and efficiency engineering can partly substitute for raw chip supply, complicating the assumption that compute limits alone will keep the frontier concentrated in a few hands.
The Reaction
Developers moved fast. Because the weights are open, hosting providers began standing up day-zero endpoints almost immediately, and independent testers started reproducing the benchmark claims rather than taking them on faith. The frontend-coding result in particular circulated widely, because it is the kind of task working engineers can sanity-check in an afternoon, and early hands-on reports were broadly positive about the model's ability to hold a long, complex task together across its million-token window.
The pricing helped the story travel. Moonshot set API rates at roughly thirty cents per million cache-hit input tokens, three dollars per million on cache misses, and fifteen dollars per million output tokens, which is aggressive for a model competing at this level. For teams that would rather self-host, the open licence removes the per-token meter entirely, at the cost of provisioning serious hardware to run a model this large.
Not everyone is celebrating without reservation. Skeptics point out that a leaderboard placement is a snapshot, not a verdict, and that the messy realities of production, latency, tooling, reliability, and support tend to favour incumbents with mature platforms. The gap between "matches on a benchmark" and "trustworthy in a shipping product" is exactly where closed providers still argue their value lies.
What Comes Next
The most immediate question is how the American frontier labs respond. When an open model closes to within a few points of the best closed systems, the pressure runs in two directions at once: it pushes closed labs to justify their prices with capabilities that cannot be easily copied, and it tempts them to release more of their own work openly to stay part of the conversation. Either way, the release resets what developers can expect to get for free.
A second question is durability. Open weights invite a global community to fine-tune, compress, and specialise a model in ways its creators never anticipated, which can extend a model's useful life well beyond its launch window. If a healthy ecosystem forms around K3, its influence may be measured less by this month's benchmark than by the derivatives, tools, and smaller distilled versions that grow out of it over the coming quarters.
There is also a policy dimension that will not resolve quickly. A capable, freely downloadable frontier model is difficult to govern through the usual levers of access control and export policy, because once the weights are public they cannot be recalled. Expect regulators on both sides of the Pacific to spend the second half of 2026 arguing about what open release means for safety, competition, and national advantage.
Closing Thoughts
Moonshot took its name from the ambition of a moonshot, and Kimi K3 fits the metaphor in a way that goes beyond branding. The company reached for a target that seemed reserved for the best-funded American labs and, by most independent measures, came remarkably close, then handed the result to anyone willing to download it. Whether that turns out to be a generous act, a competitive gambit, or both at once, it changes the terms of the debate.
The larger takeaway is that the centre of gravity in AI is shifting from raw capability toward access. For a while the interesting question was who could build the smartest model. Increasingly the interesting question is who can put a very good model in the most hands, and what the world does once that model is impossible to take back. Kimi K3 does not settle that argument, but it makes it unavoidable.
한글 요약
중국 베이징의 AI 스타트업 문샷AI가 7월 26일 초대형 언어모델 '키미 K3'의 가중치를 공개했습니다. 총 2조 8천억 개 파라미터 규모로, 지금까지 공개된 오픈웨이트 모델 중 가장 큰 규모입니다. 전체 896개 전문가 네트워크 중 토큰마다 16개만 활성화되는 전문가혼합(MoE) 구조를 채택해, 밀집형 모델보다 훨씬 낮은 비용으로 이 규모를 운영할 수 있게 설계했습니다. 100만 토큰 컨텍스트 창, 시각 이해, 상시 추론('thinking mode')을 지원하며, 수정 MIT 방식의 관대한 라이선스로 배포됩니다.
독립 평가기관 아티피셜 애널리시스 기준으로 K3는 종합 Elo 약 1,547로 앤트로픽의 클로드 페이블 5에 이어 2위, 오픈AI GPT-5.5보다 앞섰고, 프런트엔드 코딩 아레나에서는 페이블 5를 제치고 1위를 기록했습니다. 핵심 의의는 순위 자체가 아니라, 누구나 내려받을 수 있는 모델이 임대해야만 쓸 수 있는 최상위 폐쇄형 시스템에 근접했다는 점입니다. 미국의 반도체 수출 규제 속에서도 이 규모를 학습·배포했다는 사실은 희소 아키텍처와 효율화가 순수 연산력의 상당 부분을 대체할 수 있음을 시사합니다.
개발자들은 공개 직후 데이 제로 호스팅을 띄우고 벤치마크를 직접 검증하기 시작했으며, 공격적인 API 가격과 자가 호스팅 선택지가 화제를 키웠습니다. 다만 리더보드 순위와 실제 제품 안정성 사이의 간극은 여전히 폐쇄형 제공자의 강점으로 남아 있습니다. 관전 포인트는 미국 프런티어 랩들의 대응, K3를 둘러싼 파생·경량화 생태계의 형성, 그리고 한번 공개되면 회수할 수 없는 오픈웨이트 모델을 어떻게 규율할지에 대한 논의입니다. 참고: VentureBeat, Tom's Hardware, Forbes.