A 2.8 trillion parameter open model lands inside AWS's compliance perimeter

Amazon Web Services confirmed on September 18, 2026 that Kimi K3, Moonshot AI's open weight model, is now generally available on Amazon Bedrock across every AWS Region where the service operates. That single move resolves the trade off that has defined open weight adoption for two years: companies could download and customize an open model freely, but they usually lost the compliance certifications, audit logging and data residency guarantees that come standard with a closed model on a managed cloud. With K3 inside Bedrock, that trade off disappears for any team already running workloads there.

Why 2.8 trillion parameters is the headline number worth pausing on

Kimi K3 is now the largest open weight model available on a major cloud platform. According to AWS's own announcement, the model combines native vision, a 1 million token context window, and Mixture of Experts routing that activates roughly 104 billion parameters per request rather than the full 2.8 trillion. That routing detail matters practically: inference cost tracks the active slice of the network, not the total weight count, so the jump in scale does not translate one to one into a jump in serving cost. Moonshot AI also claims an approximate 2.5 times improvement in scaling efficiency over its predecessor Kimi K2, a vendor reported figure that AWS repeats without independent verification.

The detail that actually changes deployment decisions

AWS says this is the first open weight model on Bedrock to support explicit prompt caching, letting developers mark the end of a reusable prompt prefix and pay a discounted rate on the input tokens that hit that cache on subsequent calls. For production systems that repeatedly send the same large context, a code repository, a policy document, a long conversation history, that feature is the difference between an expensive habit and a sustainable one. Access runs through the Bedrock Runtime endpoint, the native Invoke and Converse APIs, and OpenAI compatible Responses and Chat Completions APIs, plus a global cross Region inference profile that AWS prices at roughly 10% below a fixed geographic profile.

What regulated industries actually get from this

Financial services, healthcare and public sector teams that cannot send documents to an unvetted endpoint can now run K3 inside the same regional boundary, identity controls and guardrails they already apply to every other Bedrock model. That is frequently the deciding factor in whether a model gets approved for production at all, independent of how well it scores on a benchmark. It also means K3 inherits Bedrock's existing content filtering and PII detection without a separate security review, the kind of institutional friction that has kept plenty of capable open models out of regulated environments.

The open questions AWS did not answer

Independent benchmarks for Kimi K3 are not yet widely available, so the model's real world performance against GPT-6 Astra, Claude Fable 5.1 or Gemini 3.8 Flash on Bedrock specific workloads remains untested outside Moonshot's own claims. AWS has also not disclosed whether it reviewed K3 for bias or security vulnerabilities beyond its standard model onboarding process, a gap worth watching given that K3 is a Chinese developed model entering a market where data sovereignty scrutiny of exactly that category of model has intensified through 2026. For AWS, adding the largest open weight model available anywhere widens Bedrock's catalog against rivals, but it also puts the platform on the front line of that scrutiny.

Sources

Amazon Web Services, Introducing Kimi K3 on Amazon Bedrock: https://aws.amazon.com/blogs/machine-learning/introducing-kimi-k3-on-amazon-bedrock Inside AI, Kimi K3 by Moonshot AI is now generally available on Amazon Bedrock: https://insideai.news/news/ai-in-business/kimi-k3-amazon-bedrock/12298/ ByteVyte, Kimi K3 on Amazon Bedrock Brings a 2.8T Open Model to AWS: https://bytevyte.com/kimi-k3-on-amazon-bedrock-brings-a-2-8t-open-model-to-aws