Moonshot AI's Kimi K3 is now generally available on Amazon Bedrock, and it turns up with a first to its name: per AWS, it is the first open weight model on the platform to support explicit prompt caching. It arrives with a 1-million-token context window, native vision and, by Moonshot's own count, 2.8 trillion parameters.
AWS AI Watch analysis
What happened
Per the AWS What's New announcement, Kimi K3 is available in every Region where Bedrock runs, via cross-Region inferencing, and sits inside the same security boundary as proprietary models. That means the same controls for access, encryption and auditing across a customer's whole model portfolio, open weights and closed alike.
The capability claims come from Moonshot AI, relayed through the announcement: Kimi K3 is the company's most capable model, the first open model to reach 2.8 trillion parameters, and roughly 2.5 times more scaling-efficient than Kimi K2. It pairs native vision with that million-token window, which the announcement pitches at long coding sessions across large repositories, multi-document analysis including scanned pages and screenshots, and extended agent workflows.
The operational headline is caching. Explicit prompt caching cuts latency and input costs when the same context is reused across calls. Until now that has been a perk of the proprietary models on Bedrock; Kimi K3 is the first open weight model to get it.
Why it matters
The open-weight race and the enterprise procurement queue have just met in the middle. A team that liked the sound of Kimi K3 but had no appetite for operating its hardware can now point agent workflows at it through a console, with AWS handling the governance plumbing. For Chinese labs, a slot on a Western hyperscaler's managed platform is becoming a familiar route to enterprise customers, and Bedrock's open weight shelf is getting crowded in the interesting way.
Caching is the economics. Long-context work is where input tokens pile up, and repeated context is what agents and repository-scale coding generate by the lorry-load. Folding an open weight model into that pricing structure makes the managed route competitive with self-hosting on cost, not just convenience.
Our read
Split the claims in two. The Bedrock facts are AWS's own and dull in the best sense: the model is there, governed, everywhere Bedrock is. The model facts, the parameter count and the scaling figure, are Moonshot's scoreboard, and independent benchmarks will settle them in due course. If you run long-context or agentic workloads, the sensible move is cheap: benchmark K3 against your current model on your worst real task and let the caching arithmetic make the argument. The low-drama route to a 2.8-trillion-parameter open model is now a console click.
What to watch
- Independent benchmark runs, now that enterprise testers can reach the model without hosting it.
- Whether other open weight models on Bedrock gain explicit prompt caching, or Kimi K3 keeps the edge to itself.
- Pricing and throughput specifics, which the launch announcement leaves to the Bedrock documentation.
- Whether Moonshot's rivals follow the same managed-distribution route to Western enterprise customers.
Discussion spark: With a frontier open weight model now behind Bedrock's governance and billing, what would still justify running open models on your own hardware: price, data control, or plain stubbornness?
Sources and evidence
- Kimi K3 by Moonshot AI is now generally available on Amazon Bedrock (18 September 2026, 15:15 UTC)
not affiliated with or endorsed by Amazon Web Services (AWS)