Things to know about Kimi K3

Kimi K3 is the next generation model of Moonshot AI — a system with an MoE of approximately 2.8 trillion with a context window of 1 million tokens, launched on July 16, 2026 on Kimi Code and Kimi app.

What is Kimi K3?

The Kimi K3 is Moonshot AI’s new flagship model, and the outstanding numbers are truly astonishing:

• 2.8 trillion parameters — the largest model ever released to market by a Chinese AI lab, and about 2.8 times larger than its predecessor K2.6
• 1 million token context window — enough to hold entire source code, book-length documents, or days of agent activity data in a single prompt
• Built-in multimodal capabilities — understand text, images, and video built from the ground up, rather than added later
• Continuous inference — K3 always thinks before responding, with inference effort level configurable via API

What can you do with Kimi K3?

• Long session automation programming. The K2.6 has demonstrated the ability to run 12-13 hours of automated programming with thousands of tool calls, and K3 expands on this with the K3 Swarm Max variant for massive parallel processing.
• Large context research and analysis. With a confirmed 1 million token context window, K3 is in the same league as leading tools for long context for full-repository code review, long document aggregation, and multi-document research.
• Programming agent backend systems. Early tester ChrissGPT positioned the K3 pre-launch as “an Opus 4.7+ programming model” that “outperforms GPT 5.5 and GPT 5.6 Terra on SOME programming assessments,” and noted that Fable 5 and GPT-5.6 Sol still dominate on Terminal-Bench 2.1, in a July 14 post. At launch, he rated the K3 at programming levels. Opus 4.8 equivalent program.
• Multi-agent coordination. K2.6 has integrated Claw Groups to coordinate heterogeneous external actors; K3 Swarm Max continues this direction with explicit support for large-scale parallel processing.

Price of Kimi K3

All previously released Kimi versions compete mainly on price. K3 does not. Here’s how it stacks up against its peers and top competitors (all prices are per 1 million tokens; input price is cache miss rate):

Who should use Kimi K3?

Based on what has been confirmed: K3 is currently suitable for long-term agent-based tasks where quality per task is more important than cost per token — long hours programming agents navigating large repositories, multi-million token document datasets queried iteratively based on cached prefixes, and multimodal inference tasks requiring images, videos, and code in the same context. Moonshot’s onboarding documentation is aimed precisely at that data, with setup guides from day one for Kimi Code, Claude Code, Codex, Cline, RooCode, OpenCode, OpenClaw, and Hermes Agent.

However, it is not the right choice today for latency-sensitive interactive products (28 tok/s), teams that need tunable sampling or lightweight inference modes (everything is locked to the max), for processes that require open weights (unpublished), and for pure cost optimization strategies (that’s what K2.6 and DeepSeek V4 are for).