Moonshot AI's Kimi K3 goes live on Amazon Bedrock

Overview

Updated 1 hour ago

Companies can now deploy Moonshot AI's Kimi K3, a 2.8-trillion-parameter open-weight model, on Amazon Bedrock with the same security controls reserved for proprietary models. It is the largest open model on the service, with native vision, a 1-million-token context window, and explicit prompt caching, an open-model first on Bedrock.

Until now, open-weight models forced a compromise: teams could customize them freely but lost the compliance certifications and security tooling of closed models. Bedrock removes that trade-off. Customer data stays inside AWS's boundary, never reaches Moonshot, and never trains the model.

The launch also draws scrutiny. AWS has not disclosed any review of Kimi K3 beyond standard onboarding. Data sovereignty and national security concerns have been raised about Chinese models hosted on US clouds.

Why it matters

Open-weight AI now has a permanent home inside enterprise-grade clouds, giving Chinese models a direct path into US workloads.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Key Indicators

2.8T

Model parameters

The largest open-weight model on Amazon Bedrock, and the first open model to reach 3T class.

Context window

Tokens of context Kimi K3 can process in a single request, supporting long coding sessions and multi-document analysis.

2.5x

Scaling efficiency vs Kimi K2

Moonshot's claimed improvement from the new Kimi Delta Attention architecture.

16/896

Active experts per token

Experts activated per request under the Stable LatentMoE framework.

Voices

Curated perspectives — historical figures and your fellow readers.

Ever wondered what historical figures would say about today's headlines?

Organizations Involved

Moonshot AI

AI Research Company

Developer of Kimi K3; in early-stage revenue-sharing talks with US cloud providers

Beijing-based developer of the Kimi family of language models, including the 2.8-trillion-parameter Kimi K3.

Amazon Web Services (AWS)

Cloud Provider

Host of Kimi K3 on Amazon Bedrock across all AWS Regions

Amazon's cloud division, now offering Kimi K3 through its Bedrock model platform with explicit prompt caching support.

Timeline

July 2026 September 2026

4 events Latest: Today

Tap a bar to jump to that date

September 2026
Kimi K3 goes generally available on Amazon Bedrock
Today Partnership

Kimi K3 becomes generally available on Amazon Bedrock across all AWS Regions via cross-Region inference profiles, with explicit prompt caching as an open-model first on the platform.
Today
July 2026
Full Kimi K3 weights released
Product Launch

Moonshot releases complete Kimi K3 model weights under the Kimi K3 License, letting other organizations host and modify the model.
July 27th, 2026
Demand overwhelms Moonshot's capacity
Operational

Moonshot pauses new subscriptions to its hosted Kimi K3 service three days after launch as demand exceeds available computing power.
July 20th, 2026
Moonshot AI unveils Kimi K3
Product Launch

Moonshot AI reveals Kimi K3, a 2.8-trillion-parameter open-weight model with native vision and a 1-million-token context window.
July 17th, 2026

Scenarios

Moonshot lands US cloud revenue-sharing deals

Possible Resolves by Q2 2027

Discussed by: Bloomberg Intelligence analysts cited via Quartz; reporting by finance.biggo.com on the negotiations

Moonshot is seeking up to 30% of revenue from Kimi K3 services hosted on Microsoft Azure, Amazon Web Services, and Google Cloud, according to people familiar with the talks. AWS already hosts the model on Bedrock. Whether Microsoft or Google follow depends on unresolved questions about revenue split, data access, and how token usage is audited for billing.

Microsoft Azure or Google Cloud follows AWS with Kimi K3 hosting

Possible Resolves by Q1 2027

Discussed by: finance.biggo.com reporting on early-stage negotiations between Moonshot and Microsoft, Amazon, and Google

Moonshot has held preliminary talks with Microsoft and Google about hosting Kimi K3 on Azure Foundry and Vertex AI, but no agreements are guaranteed. If either cloud announces general availability, the model reaches every major US hyperscaler. If both decline, AWS becomes Kimi K3's only major US distribution channel.

US regulators impose restrictions on Chinese open-weight models

Unlikely Resolves by Q3 2027

Discussed by: insideai.news raised scrutiny questions; Noah Intelligence highlighted the geopolitical divide in cloud deployments

Hosting a Chinese-developed model inside US cloud infrastructure may draw regulatory attention. AWS has not disclosed any special review of Kimi K3 beyond standard model onboarding. New Commerce Department or CFIUS guidance targeting Chinese open-weight models in US clouds could reshape this arrangement and affect how other Chinese models reach US enterprises.

Historical Context

3 moments from history that rhyme with this story — and how they unfolded.

July 2023

Meta Llama 2 on AWS Bedrock (2023)

Meta released Llama 2 as an open-weight model and made it available through AWS Bedrock and Microsoft Azure. It was among the first major deployments of an open-weight model on enterprise cloud platforms with security controls attached.

Then

Open-weight models gained credibility for enterprise use, with major clouds providing hosting and governance.

Now

Set the pattern of open-weight models distributed through hyperscalers, later followed by Mistral, DeepSeek, and others.

Why this matters now

Kimi K3 extends this pattern at vastly larger scale and with a Chinese developer, testing how far open-weight distribution can go when geopolitical scrutiny is higher.

February 2024

Mistral on Microsoft Azure (2024)

French AI startup Mistral signed a deal with Microsoft to make its models available on Azure, with a $16 million investment from Microsoft. The deal gave Mistral global distribution while keeping its models open-weight.

Then

Mistral gained enterprise distribution through a US hyperscaler, though the investment drew European regulatory review.

Now

Showed that non-US open-weight model makers can reach enterprise markets through US clouds, but not without scrutiny.

Why this matters now

Kimi K3 is a similar arrangement with a Chinese developer, where the investment stakes, regulatory scrutiny, and geopolitical tension are all higher.

January 2025

DeepSeek R1 (2025)

Chinese lab DeepSeek released an open-weight reasoning model that matched OpenAI's top-tier reasoning model on several benchmarks at a fraction of the cost. The release triggered a global selloff in US AI stocks, with Nvidia losing hundreds of billions in market value in one session.

Then

US lawmakers introduced bills to restrict Chinese AI; US tech companies scrambled to explain the competitive gap.

Now

Established that Chinese labs can produce frontier open-weight models, reshaping US-China AI competition.

Why this matters now

Kimi K3 is the next major Chinese open-weight frontier model, but unlike DeepSeek R1, it now has enterprise distribution inside a US hyperscaler rather than relying on direct downloads.