Amazon Nova 2 Lite

Amazon Nova 2 Lite launched on 2 December 2025 at re:Invent and marks the point where Amazon stopped selling a wide ladder of Nova models and started selling one. It takes text, images, video and documents, returns text, holds a one-million-token context window and can generate up to 65,536 tokens in a single response. The headline feature is extended thinking with a dial: three intensity levels — low, medium and high — that let the caller trade latency against accuracy per request rather than per model. Amazon also built two tools directly into the model instead of leaving them to the integrator: web grounding, which pulls live information from the web, and a code interpreter that executes Python for calculations. Remote MCP tools are supported, which matters for anyone assembling agents out of external services. The pricing is the argument. At USD 0.30 per million input tokens and USD 2.50 per million output tokens, Nova 2 Lite sits below the comparable efficient tiers from Google, OpenAI and Anthropic, and Amazon leans on price-performance rather than benchmark leadership in its own materials. Independent aggregations place it around 64 on SWE-bench Verified and around 62 on MMMU-Pro — behind the strongest small models from its rivals, ahead of some, and cheap enough that the gap is a deliberate trade. Context explains the name as much as the specification. The first Nova generation — Premier, Pro, Lite, Micro, Canvas, Reel and Sonic — is being wound down through September 2026, and a Nova 2 Pro announced alongside this model remained in preview, available only to customers of Amazon Nova Forge, the service that lets organisations train their own frontier models on Nova foundations. For general Bedrock users in 2026, Nova 2 Lite is effectively the Nova family.

▸Videos

No videos yet.

▸Reviews