Moonshot AI Unveils Kimi K3 Open-Source Model with 2.8 Trillion Parameters, Raises $3.5 Billion in Series F

Moonshot AI released the Kimi K3 model weights, a technical report and three key infrastructure technologies underpinning the model's training — MoonEP, FlashKDA and AgentEnv — on July 27, 2026. Kimi K3 has a total parameter count of 2.8 trillion, adopts a mixture-of-experts architecture, features native visual understanding capability and supports a 1-million-token context window. The model is built on the self-developed Kimi Delta Attention hybrid linear attention mechanism and Attention Residuals technology, achieving a roughly 2.5× improvement in scaling efficiency over the previous generation. The open-source release uses a modified MIT license, enabling developers worldwide to download the weights for local deployment and secondary development.

In benchmark tests, Kimi K3 posted notable results. It scored 57.1 on the Artificial Analysis Intelligence Index v4.1, ranking fourth globally and the highest among open-source models. In the Frontend Code Arena, a large-model coding evaluation system, Kimi K3 topped the leaderboard with an Elo of 1,679, becoming the first open-source model to surpass closed-source competitors and claim the top spot. The technical report also noted that its overall performance remains slightly behind Claude Fable 5 and GPT-5.6 Sol.

Moonshot AI completed a Series F funding round on July 29, raising over US$3.5 billion at a pre-money valuation of US$50 billion as part of its pre-IPO financing.

Following the open-source release, the domestic computing ecosystem quickly adapted. Huawei Ascend CANN announced that the full Ascend 950 series and Atlas A3 products support Kimi K3 deployment, while Qujing Technology completed inference adaptation on the Ascend Atlas A3 using SGLang. Alibaba Cloud's Zhenwu M890 super-node instance achieved same-day adaptation, and the Qianwen AI platform and Alibaba Cloud Bailian will offer Kimi K3 model APIs. Hygon completed full-process adaptation and performance validation on its DCU platform. Moore Threads adapted the model on its MTT S5000 intelligent computing card, and Runjian announced that Kimi K3 has been added to its computing service platform "Rundao Xing Shan".

Overseas, AI infrastructure providers including Nebius, Baseten and Fireworks announced same-day adaptation of Kimi K3. Nebius described the model as the first open-weight model to reach frontier performance levels. Well-known AI coding company Cursor has integrated Kimi K3, and Cognition, the developer of the digital worker Devin, announced that Kimi K3 is now accessible through its desktop client and command-line tools. The model topped the Hugging Face trending chart within 30 minutes with over 4,000 likes, setting a record for the fastest release growth on the platform.

The technical report disclosed several architectural details: the model mixes KDA and Gated MLA at a 3:1 ratio for efficient long-context modeling and enhances cross-layer information flow through block-level attention residuals. It uses the Stable LatentMoE mechanism, activating 16 out of 896 routed experts per token, with SiTU-GLU and Quantile Balancing ensuring training stability. The vision encoder MoonViT-V2 was trained using next-token prediction. Among the three open-sourced infrastructure technologies, MoonEP is a high-performance expert parallel communication library for large-scale MoE, FlashKDA delivers a 1.72× to 2.22× Prefill speedup over baseline on Nvidia H20 GPUs, and AgentEnv is a sandbox system supporting large-scale multi-agent parallel training.

Additionally, the model was used to independently develop MiniTriton, a compact GPU compiler, and to complete a 45nm inference chip prototype design within 48 hours using open-source EDA tools.

API pricing for Kimi K3 is set at RMB 2 per 1 million tokens for input with cache hit, RMB 20 for input with cache miss, and RMB 100 for output, with a context window of approximately 1.0486 million tokens.