Kimi K3 brings open-source AI near closed-model frontiers

For an AI agent handling hundreds of thousands of tokens, Kimi K3’s headline is not only its size. Moonshot AI’s model combines 2.8 trillion parameters with a 1-million-token context window, and a BOCOM International analysis places it near the capability of leading closed-source systems.

The comparison is unusually close. K3 scores around 57 on third-party composite intelligence indexes such as Artificial Analysis, versus around 59 for GPT-5.6 Sol and around 60 for Claude Opus 5/Fable 5, according to the report. The gap comes with a lower estimated price: roughly $0.86 per task, about 70% of GPT-5.6 Sol and 30–40% of Anthropic’s flagship models.

The architecture is built to make long context less expensive to process. KDA and gated MLA handle million-token sequences, while attention residuals allow selective retrieval across layers. A stable latent Mixture-of-Experts system — a design that routes each token through only part of a larger model — mixes 896 experts, with 16 active at a time and 104.2 billion parameters activated per token.

Moonshot’s systems work is part of the explanation. Its MoonEP infrastructure uses redundant copies of experts and balances token distribution across GPUs, addressing the uneven workloads that can leave some chips overloaded while others sit idle. The report says this pushes cluster utilization toward its physical limit and cuts training costs, helping make a model of K3’s scale manageable.

So what changes in practice? If the report’s cost and capability estimates hold up, companies could run more capable agent-based applications without paying the prices associated with the leading closed models. That matters most for tasks that repeatedly search large bodies of information or maintain very long working contexts. But the figures remain those cited by BOCOM International and third-party indexes; no independent deployment test is presented here.

The commercial bet is already substantial. Moonshot closed a $3.5 billion Series F, lifting its valuation 75% to $35–50 billion since May. A pre-IPO valuation of $50 billion is expected to rest partly on the possibility that Kimi ARR could reach about $1 billion in the next 6–12 months. The report identifies compute access and the ability to keep iterating — rather than parameter count alone — as the harder test ahead.