Amazon Nova Micro

Amazon Nova Micro launched on 5 December 2024 alongside Lite and Pro, and it is the odd one out of the family. Every other Nova understanding model takes images and video; Micro takes text and nothing else, has no document support at all, and returns text. Its context window is 128,000 tokens against 300,000 for Lite and Pro, and it can return up to 10,000 tokens. What it buys with those restrictions is the lowest latency and the lowest price in the range — roughly half the input cost of Nova Lite, which is itself the cheap tier. That trade has a clear shape of use. Micro is aimed at high-volume text work where the model is called constantly and each call has to be nearly free: classification, routing, extraction from plain text, short chat turns, the small decisions inside a larger agent. Amazon supports it with the full Bedrock toolchain — fine-tuning, provisioned throughput, batch inference, Agents, Knowledge Bases, prompt flows — and, unusually, Guardrails and Evaluations work on it without the text-only caveat the multimodal Novas carry, because Micro is text-only to begin with. It is a distillation student of both Premier and Pro. It speaks the same 200-plus languages as the rest of the family, optimised for fifteen. As of August 2026 Micro is Active on Amazon Bedrock and does not appear in the lifecycle table of legacy or retiring models, unlike Nova Premier, Sonic, Canvas and Reel, which all shut down in September 2026. It runs in Northern Virginia, Tokyo, Sydney, London and GovCloud, with cross-region inference across the US, Europe and Asia Pacific.

▸Videos

No videos yet.

▸Reviews