Moonshot AI has released Kimi K3, a 2.8 trillion parameter model that the company calls the "world's first open 3T-class model". According to the official docs, K3 is a mixture-of-experts model that activates 16 of 896 experts per token, with native vision capabilities and a 1M token context window, and two new attention mechanisms. Moonshot claims these architectural changes yield roughly 2.5× better scaling efficiency than K2, and the model targets long horizon coding, knowledge work and complex reasoning. It's available now through the Kimi app, Kimi Code, and the API, with full weights promised by July 27.
The company's own blog honestly admits K3 "still trails" Claude Fable 5 and GPT 5.6 Sol, the two proprietary models at the top of the market. But below that line, K3 wins or places second in most of the coding and agentic evaluations against Claude Opus 4.8, GPT 5.5, and Chinese rival GLM-5.2, and independent testing by Artificial Analysis largely confirms the picture, placing it around Opus 4.8 tier. An Open Source model outperforming last generation's Western flagships, weeks after their successors shipped, compresses the lag between proprietary and open frontier to something like a single release cycle. Remember that the gap used to be measured in a year or more!
Then there's the price, which cuts in two directions. At $3 per million input tokens and $15 per million output, K3 is way cheaper than the models it's chasing. Fortune notes how Fable 5 charges $50 for the same output volume, but K3 is also far more expensive than what Chinese labs have accustomed us to. For context, GLM-5.2 costs $4.40 per million output tokens and DeepSeek V4 only $0.87. Chinese AI is no longer competing purely on being cheap; Moonshot is betting there's a market for a near-frontier model at a mid tier Western price, which is a very different proposition than the "race to the bottom" dynamic DeepSeek triggered in early 2025.
The Open Source label deserves some scrutiny too. Until the weights actually land on July 27, K3 is a hosted model like any other (in fact, Artificial Analysis is classifying it as proprietary for now), and even once released, a 2.8T parameter model that Moonshot recommends deploying on 64 or more accelerators is not something we can run at home precisely. Open weights, at this scale, are less about individual access, and more about giving enterprises a frontier foundation they can inspect and serve without paying Anthropic or OpenAI margins. Still important though, for pricing pressure, for research, for countries and for companies wary of dependence on American APIs, but it's a different kind of openness than what made earlier Kimi and DeepSeek releases community darlings.
For the big western labs, their moat is now the distance between Fable 5/GPT 5.6 Sol and K3, sustained release after release, while their pricing power rests on a premium that an open alternative keeps undercutting from below. Moonshot itself lists actual, important limitations (instability outside its own harness and a tendency to improvise beyond user intent, and of course the experience gap with the Frontier) so the top tier isn't seriously threatened today. But, when the second-best tier is open and cheaper, we can only wonder how much anyone is willing to pay for the lead.