Kimi K3 Just Redrew the Open-Weight Map — 2.8T Parameters, 1M Context, and a Price That Makes GPT-5 Look Lazy

Kimi K3 Just Redrew the Open-Weight Map — 2.8 Trillion Parameters, 1M Context, and a Price That Makes GPT-5 Look Lazy

Late Thursday night, Beijing-based Moonshot AI pushed the full weights of Kimi K3 to Hugging Face under a custom license and called it something the rest of the open-weight field has been afraid to claim: "open frontier intelligence." The number that did the talking is 2.8 trillion parameters in a sparse Stable LatentMoE design that activates 16 of 896 routed experts per token. That makes K3 the largest openly available language model on the planet, well ahead of the previous open-weight champion, DeepSeek V4, and roughly an order of magnitude bigger than any open model coming out of a Western lab. It is not a research artifact. It is shipping, it has a real API, and it is priced in a way that forces a fresh conversation about what "frontier" actually costs.

What the architecture actually is. K3 is a sparse mixture-of-experts with three pieces that matter to anyone running it. The routing: 896 experts, only 16 active per token, so the effective compute per forward pass is a fraction of the headline parameter count. Moonshot has replaced standard multi-head attention with Kimi Delta Attention, a linear-attention variant that scales more gracefully as context length grows, paired with Attention Residuals that recover some of the quality loss that linear attention usually inflicts. K3 is multimodal from the first layer — native visual understanding is built in, not bolted on — and the context window is one million tokens. The architectural story is that Moonshot is actually trying to make long windows economically usable, not just spec-sheet-impressive.

The benchmark picture is honest. On the Artificial Analysis Intelligence Index, K3 scores 57. OpenAI's GPT-5.6 Sol sits at 59. That is a one-percent gap to the current closed-source frontier, and K3 gets there with a model anyone can download. On BenchLM it lands at 79.9/100 and ranks number five of 216 models tested, with its strongest category being Agentic at number four. On the SWE Marathon leaderboard — the long-horizon software engineering benchmark that has started to displace SWE-Bench for real engineering work — K3 is currently number one. None of these are marketing slides. The interesting thing is not that K3 is best in class. It is that K3 is in the conversation at all, at open-weight prices.

The price is the story nobody is talking about enough. K3's API is $3 per million input tokens, $15 per million output, with cached input at $0.30 per million. For the agent loops that chew through hundreds of thousands of tokens per task, the cache-hit price is the one that actually matters, and it is materially cheaper than what OpenAI and Anthropic charge for comparable cached-input access. The combined read is simple: an open-weight model one percentage point behind the closed frontier on aggregate intelligence, cheaper on inference, and downloadable for self-hosting. For any team that treated frontier model access as a pay-OpenAI-or-die decision, that table just changed.

The license is the asterisk. Moonshot is calling K3 an open-weight model, and that is technically true. The license is a custom MIT-derived document that lets you download, run, modify, fine-tune, distribute, and deploy commercially at no license cost. It is not open source in the OSI sense — there are usage restrictions, a brand rule around the Kimi name, and a roughly $20M MaaS revenue gate that triggers extra obligations once you start reselling at scale. For most developers and small teams, none of this matters. For anyone building a commercial product on top of K3, read the license before you ship. "Open weights" is not the same as "open source," and being honest about that distinction is the difference between a sustainable product and a takedown notice.

Why this matters beyond the headline. The pattern of the last eighteen months has been: closed frontier lab ships a new model, the open-weight community plays catch-up six months later with a smaller but usable version. K3 breaks that pattern in two ways. First, the gap between closed and open is now small enough to be measured in single-digit percentages on aggregate intelligence, not in capability categories. Second, the moat that closed labs leaned on — being the only place where frontier capability exists — is gone for the parts of the frontier that actually matter to production systems: reasoning, coding, agentic tool use, long-context document work. Every procurement conversation about AI infrastructure that happened before this week is stale. Every one that happens after it has a third option on the table that did not exist seven days ago.