Kimi K3 by Moonshot AI now generally available on Amazon Bedrock

AWS made Moonshot AI's Kimi K3 generally available on Amazon Bedrock, positioning it as an open-weight option for coding and knowledge work. The model offers native vision, a 1-million-token context window, and explicit prompt caching, and Moonshot describes it as the first open model to reach 2.8 trillion parameters — a scale claim that, if accurate, makes it one of the largest openly available models on the market.

The Bedrock GA is significant because it puts a frontier-scale Chinese open-weight model onto AWS's managed inference platform alongside OpenAI's GPT-6 Astra and Anthropic's Claude Fable 5.1, giving enterprises a single procurement path across closed and open frontier models. For AWS, hosting the largest open model reinforces its 'model-neutral platform' pitch against Azure's OpenAI-centric and Google's Gemini-centric clouds.

The 1M-token context lands in the same week as Alibaba's Qwen3.8-Omni-Flash, both at the top of the long-context tier, making 1M tokens the new competitive baseline for serious coding and document workloads. The 2.8-trillion-parameter figure drew developer attention as a marker of how aggressively Chinese labs are scaling open models — a theme reinforced by reports this week that six or seven Chinese labs are targeting 10–40 trillion-parameter models within two years.

The practical questions for adopters: inference cost at that parameter scale even with MoE-style activation, how prompt caching offsets the long-context expense, and whether an open-weight model of this size is genuinely self-hostable outside a hyperscaler. For most enterprises the answer is that Bedrock hosting, not the open weights themselves, is the realistic access path.