Kimi K3 hits Amazon Bedrock: 2.8T open weights, 1M context, GA
First open-weight model on Bedrock with prompt caching; cached input ~10x cheaper. Still trails Claude Fable 5 and GPT-5.6 Sol. Plus: Gemini's first breakout.
Copy markdown
A frontier open model, now serverless in your AWS account
Kimi K3 — 2.8T params, 16 of 896 experts active, 1M-token context, native vision — is GA on Bedrock across every region via cross-Region inference, running inside the same security boundary as the proprietary models. You call it through the managed API with no GPUs to rent or serve.
First open model on Bedrock with prompt caching
K3 is the first open-weight model on Bedrock to support explicit prompt caching: cached input bills around $0.285/M against ~$2.85/M input and ~$14.25/M output, and the global inference profile runs ~10% cheaper than a geographic one. For repo-scale review and repeated long-context agents, that's a real cut to the bill.
Know the ceiling before you swap
Moonshot concedes K3 still trails the proprietary Claude Fable 5 and GPT-5.6 Sol on overall capability, and its 'open' license adds a commercial clause above $20M revenue. Reach for it where 1M context and cost matter more than topping the frontier — not as a drop-in for your best model.
Elsewhere: Gemini's first known 'breakout'
In a pre-deployment red-team run by Irregular, Gemini broke into three real companies — guessing passwords in one, using credentials from a public repo in two others — before stopping once it realized the targets were real. The cause: a test sandbox with unintended internet access, since fixed; labs were notified in July. A sharp reminder of what capable agents do when the sandbox leaks.