Kimi K3 pushes open-weight AI toward frontier performance, with higher prices attached

Kimi is launching K3 as a multimodal open-weight model with 2.8 trillion parameters and a one million-token context window. The model processes images and video natively and is aimed at long-running programming tasks, knowledge work, and complex reasoning. Full model weights are scheduled for release by July 27, making availability a key point to watch for developers and teams evaluating open models.

In Kimi’s own benchmarks, K3 trails Claude Fable 5 and GPT-5.6 Sol but beats other tested systems, including Opus 4.8 and GLM 5.2, in several comparisons. Across 35 tests, it took first place about seven times and was second or third in most others. Independent testing from Artificial Analysis broadly supports the performance picture, while also flagging that K3’s hallucination rate rose compared with K2.6.

Kimi positions K3 for full development projects rather than short code snippets. The model is designed to analyze large codebases, coordinate terminal tools, and maintain focus across many steps with minimal human oversight. Its “Vision in the Loop” workflow lets it inspect screen captures, modify code, and check visible output, which Kimi connects to use cases such as game development, UI design, and CAD.

K3 is priced at $3 per million input tokens without a cache hit, $0.30 with a cache hit, and $15 per million output tokens. That is significantly above K2.6’s listed prices and suggests Chinese providers are no longer reserving frontier models for ultra-low-cost API tiers. Artificial Analysis puts K3 at about $0.94 per task on its Intelligence Index, close to GPT-5.6 Sol and below Opus 4.8, but still above open-weight peers such as GLM-5.2 and DeepSeek V4 Pro.

K3 is already available through Kimi.com, mobile apps for iOS, Android, and HarmonyOS, the Kimi Work desktop client, and Kimi Code. It is also listed on OpenRouter under the identifier “moonshotai/kimi-k3,” though the article says it is currently served there only through Moonshot itself. Kimi also plans a hosted agent platform for isolated environments and runtimes for long-running tasks.