Kimi K3 brings 2.8 trillion open weights to Bedrock
Moonshot AI's 2.8-trillion-parameter Kimi K3 is now generally available on Amazon Bedrock, with a one-million-token context window and native vision.
Amazon Bedrock has added Kimi K3, the model Moonshot AI calls its most capable yet and the first open model to reach 2.8 trillion parameters. AWS published the general availability notice on 18 September, placing the model alongside the proprietary systems Bedrock already serves.
What K3 brings to Bedrock
The headline numbers come from Moonshot AI itself: 2.8 trillion parameters, native vision, and a one-million-token context window. AWS pitches the combination at long coding sessions across large repositories, multi-document analysis that includes scanned pages and screenshots, and agent workflows that run for a long time without losing the thread.
Moonshot AI reports roughly 2.5 times the scaling efficiency of Kimi K2, the model it replaces at the top of the family. On Bedrock, K3 runs inside the same security boundary as the closed models, with the same controls for access, encryption and auditing, and it is the first open weight model on the platform to support explicit prompt caching, which cuts latency and input costs when context is reused between calls.
Availability is broad but indirect. AWS says the model can be reached in every region where Bedrock operates through cross-Region inferencing, so requests are routed to capacity rather than pinned to the region you call from.
Why an open model on a managed platform matters
Open weights usually mean downloading a very large file and finding somewhere to run it. Putting a 2.8-trillion-parameter model behind a managed API changes who can use it: the weights are still publishable, but the infrastructure, the billing and the compliance paperwork stay with the cloud provider. For companies that want the flexibility of an open license without owning a cluster of GPUs, that is the trade they have been waiting for.
Our opinion
The interesting part of this announcement is not the parameter count, which is a marketing number as much as a technical one, but the prompt caching. Kimi K3's selling point is a million-token context, and context that long is only affordable if you stop paying to re-send it on every call; being the first open weight model on Bedrock to get that treatment says more about where the economics of long-context work are heading than any benchmark figure. The second interesting thing is the phrase 'according to Moonshot AI'. AWS is careful to attribute the 2.8 trillion and the 2.5x scaling claim to the model's maker rather than measuring them itself, which is the honest way to write a launch note and a reminder to read the claims accordingly. Two years ago a Chinese open-weight release arriving on the biggest Western cloud on the same day would have been a story about geopolitics. Now it is a line in a product changelog, and that normalisation is the real news.