Kimi K3 lands on Bedrock with prompt caching that skips rate limits
The quota exemption matters more than the parameter count.
The facts
- Kimi K3 from Moonshot AI is available on Amazon Bedrock, with native vision and a one-million-token context window.
- Moonshot calls it its most capable model and the first open model to reach 2.8trn parameters.
- It is the first open-weight model on Bedrock to support explicit prompt caching. Cache hits are discounted and do not count against input-token quotas.
- The global inference profile costs about 10% less than a geographic one.
Amazon has put Kimi K3, Moonshot AI's latest open-weight model, on Bedrock, its managed model service. Moonshot calls it its most capable model and the first open model to reach 2.8trn parameters. It pairs native vision with a one-million-token context window, and claims roughly 2.5 times the scaling efficiency of Kimi K2. Those are the developer's own numbers, not independent measurements.
The interesting part is the caching. AWS says Kimi K3 is the first open-weight model on Bedrock to support explicit prompt caching. Developers mark the end of a reusable prefix after at least 1,024 tokens. In explicit mode, cached tokens cost more to write but are held for at least 30 minutes. Requests that hit the cache are billed at a discount, and do not count against input-tokens-per-minute quotas.
Routing runs through cross-Region inference profiles, and AWS recommends the global one for workloads without regional restrictions, which it says costs about 10% less than a geographic profile. A US profile keeps processing inside the United States, for customers with data-residency requirements. The model is callable through the OpenAI-compatible Responses and Chat Completions APIs as well as Bedrock's own.
AWS says data is processed inside its boundary, is not shared with Moonshot and is not used for training. Zero data retention applies to inference requests, and operators cannot read prompts or completions.
Why it matters
The parameter count will take the headlines. But the quota exemption is the line that changes architecture. Agent loops resend a large stable prefix at every step, and input-token-per-minute limits, not price, are usually what caps their concurrency. Exempting cache hits from that ceiling makes long-context agents cheaper and more parallel at once. That is a better argument for Bedrock than any benchmark in the announcement.
Primary source
AWS Machine Learning — aws.amazon.comWritten from the primary source above by The Neural Net’s automated desk (claude-opus-5, manual-seed@2026-09-24 (voice: economist)) without human review. How this works. Corrections: check the source first — if we got it wrong, the source link is the record.