Moonshot AI's Kimi K3 — First 2.8T-Parameter Open Model — Goes GA on Amazon Bedrock

Kimi K3's general availability on Amazon Bedrock puts frontier open-model scale into a managed hyperscaler environment: at 2.8 trillion parameters, AWS bills it as the first open model to cross that threshold, edging past even xAI's reported 2.5T Grok 4.8 target. Native vision and a 1-million-token context window position it for long-context coding and knowledge workflows where enterprises want capability without wrangling their own infrastructure.

The operative feature for cost-conscious teams is explicit prompt caching — letting developers cache and reuse large context blocks to cut both latency and input token costs on repeated calls, which matters enormously at 1M-token context sizes. Delivering this as a Bedrock-managed endpoint means enterprises get the scale of an open model with AWS's security, autoscaling and compliance wrapper.

Strategically, offering a Chinese-origin (Moonshot AI) open model on Bedrock underscores AWS's model-neutral marketplace posture — the same week it also shipped a reworked AgentCore runtime and ran open-weight models in its European Sovereign Cloud. It intensifies the open-vs-closed and US-vs-China dynamics running through the week, alongside Alibaba's Qwen releases and the FBI's accusations against Qwen.

The caveats are the familiar ones for very large open models: 2.8T parameters is enormous to serve, so real-world cost and latency depend heavily on how well prompt caching and Bedrock's optimizations work in practice. Data-governance questions around a Chinese-origin model on US cloud infrastructure will also follow enterprise buyers. Watch adoption metrics and independent benchmarks versus closed frontier models.