Chinese artificial intelligence start-up Moonshot AI made its flagship Kimi K3 model available on Amazon Bedrock on 18 September 2026, an arrangement that appears to trigger the commercial licence Moonshot adopted in July 2026, which requires a paid arrangement from any firm that hosts the model as a service and generates more than US$20 million in aggregate annual revenue, the South China Morning Post reported.
Key points
- Kimi K3 launched on Amazon Bedrock on 18 September 2026, joining open-weight models from DeepSeek, MiniMax and Alibaba already on the platform.
- Moonshot’s updated licence mandates paid terms for any firm that hosts the model as a service and exceeds US$20 million in aggregate annual revenue.
- AWS reported its AI business surpassed a US$25 billion annual revenue run rate in July 2026.
- The model carries 2.8 trillion parameters, a 1-million-token context window, and native vision capabilities.
- Moonshot was seeking up to 30 per cent of revenue in talks with Microsoft and Google, per Reuters.
Licence terms and AWS’s US$25 billion AI run rate
Moonshot revised its commercial licence in July 2026 alongside the Kimi K3 release. The terms state that any firm or its affiliates hosting the model as a service must enter a paid arrangement once aggregate annual revenue exceeds US$20 million. AWS reported in July 2026 that its AI business had passed a US$25 billion annual revenue run rate. Neither Moonshot nor AWS has disclosed the financial terms of the Bedrock agreement. An Amazon representative referred the South China Morning Post to the launch announcement, which did not address revenue-sharing details, and Moonshot did not immediately respond to the newspaper’s request for comment.
Model specifications and architecture
According to Moonshot AI, Kimi K3 is the first open model to reach 2.8 trillion parameters, AWS’s launch announcement says. It combines native vision capabilities with a 1-million-token context window and delivers an approximate 2.5x improvement in scaling efficiency over its predecessor, Kimi K2. The model uses a mixture-of-experts design called Stable LatentMoE, activating 16 of 896 experts, and applies quantisation-aware training from the supervised fine-tuning stage using MXFP4 weights with MXFP8 activations. Moonshot attributes the efficiency gain to these structural changes together with refined training and data recipes, as reported by Unite.AI.
Prompt caching and platform integration
AWS says Kimi K3 is the first open-weight model on Amazon Bedrock to support explicit prompt caching, a feature aimed at long-running coding and knowledge workflows that repeatedly resend stable context such as repository instructions or reference documents. Developers can mark the end of a reusable prompt prefix after at least 1,024 tokens using a prompt_cache_breakpoint marker. Tokens written to cache are billed at a higher rate but retained for at least 30 minutes — subsequent cache hits are billed at a discounted input rate and do not count against input-tokens-per-minute quotas. The model is accessible via the global cross-region inference profile global.moonshotai.kimi-k3, which AWS says costs approximately 10 per cent less than a geographic profile, and the US geographic profile us.moonshotai.kimi-k3 for data residency requirements.
Data handling and compliance posture
AWS states that customer data sent to Kimi K3 is processed within the AWS data boundary, is not shared with Moonshot AI, and is not used to train the underlying model. Zero data retention is always enabled for inference requests, and zero operator access prevents AWS operators from accessing prompts and completions during inference. AWS says the same protections apply to every open-weight model on Bedrock.
Chinese labs’ overseas revenue push
The Bedrock placement comes as Chinese AI developers seek to close a monetisation gap with US rivals. Seven major Chinese AI developers combined generate only about 10 per cent of the revenue reported for OpenAI and Anthropic, according to Rhodium Group. MiniMax, which also hosts models on Bedrock, Microsoft Azure and Google Cloud, reported its business-to-business revenue surged more than 700 per cent in the first half of 2026, with overseas markets accounting for about 60 per cent. Alibaba’s Qwen 3.8-Max licence, released last month, requires a separate deal from any model-as-a-service or AI work assistant business whose aggregate revenue exceeds US$50 million over any consecutive 12-month period, according to the South China Morning Post, which Alibaba owns. Moonshot was in talks with Microsoft and Google over revenue-sharing agreements for Kimi K3, seeking as much as 30 per cent of revenue, Reuters reported in August 2026.
Limitations acknowledged by developer
Moonshot AI states that Kimi K3’s overall performance still trails proprietary models including Claude Fable 5 and GPT 5.6 Sol. The company lists limitations: the model was trained in a preserved thinking history mode, so generation quality can become unstable if an agent harness fails to pass back historical thinking content, and its emphasis on long-horizon tasks can lead it to make unexpected decisions when user instructions are ambiguous, which may require more explicit behavioural constraints in the system prompt.
