Moonshot AI has launched Kimi K3, a new flagship it is calling open frontier intelligence, and by the company's numbers it is the largest open model released to date, a 2.8 trillion parameter system that runs a full million tokens of context and takes text, images, and audio natively. The model went live through Moonshot's API today, with the open weights and a technical report set to follow shortly, and it lands as the most ambitious entry yet in a year of Chinese labs closing the gap with the Western frontier.
The headline is not just the size but how Moonshot made a model this large practical to run at long context. K3 introduces a new attention design the company calls Kimi Delta Attention, which it says delivers up to 6.3 times faster decoding when the context stretches into the millions of tokens, paired with a technique it calls Attention Residuals that it open sourced earlier this year. Long context has been easy to advertise and painful to actually serve, because the memory and compute cost climbs steeply as the window grows, and Moonshot is pitching K3 as a model that can hold a million tokens without the usual collapse in speed.
K3 is natively multimodal, with visual understanding built into the model rather than bolted on, and Moonshot is releasing it in the open, continuing the strategy that made its earlier Kimi models popular with developers who wanted frontier behavior they could run and adapt themselves. For now the model is reachable through Moonshot's API and early third party providers, priced at 3 dollars per million input tokens and 15 dollars per million output tokens across the full million token window, with the downloadable weights and the full architecture writeup promised soon.
On performance, Moonshot says K3 reaches frontier level and, on its own internal evaluations, ranks just behind the strongest proprietary systems, including OpenAI's GPT-5.6 Sol, while beating a number of them on specific coding and agentic tasks. Those are the company's own figures, and independent third party benchmarks are still pending, so the real test will come as outside researchers put it through its paces. What is not in doubt is the ambition, an open model swinging at the very top of the leaderboard rather than settling for good enough.
The significance is where the open frontier now sits. For most of the past two years the assumption was that the best models would be closed and American, with open and often Chinese models trailing a step behind, and K3 is the clearest attempt yet to erase that gap, a model that is at once the biggest open release ever, one of the longest in context, and, if the weights land as promised, something any company can run and adapt on its own hardware. Independent benchmarks will decide how close to the very top it really is, but the direction is unmistakable, the open side of AI is no longer content to follow, and the largest model in the world you can soon download for yourself now comes from Moonshot.