The short version
On December 2, 2025 at AWS re:Invent in Las Vegas, Amazon unveiled the Nova 2 model family and a new build service called Nova Forge. Nova 2 arrives in tiers—Lite (GA), Pro (preview), Omni (preview) and Sonic (GA)—with long-context reasoning, built‑in tools, and multimodal options. Nova Forge lets enterprises “open up” the training pipeline by starting from early Nova checkpoints to create domain‑expert models. Nova 2, Nova 2 Omni, Nova 2 Sonic, and Nova Forge are live or in preview now.
In parallel, AWS deepened its NVIDIA partnership: it will adopt NVLink Fusion—a chiplet‑based, rack‑scale interconnect platform—for future Trainium4 accelerators (and even Graviton CPUs and the Nitro System), tying AWS custom silicon into NVIDIA’s NVLink 6 fabric and MGX rack architecture. The move aims to shorten time‑to‑market for cloud‑scale AI infrastructure.
Nova 2, at a glance
Amazon’s second‑generation models focus on controllable reasoning, longer context, and built‑in “do‑more” tools.
- Extended thinking with selectable budgets (low/medium/high) for step‑by‑step reasoning and task decomposition.
- Built‑in tools such as a code interpreter and web grounding to fetch and validate current information.
- A one‑million‑token context window across Nova 2 variants, with multimodal inputs varying by model.
Nova 2 models overview
| Model | Inputs | Outputs | Highlights | Context | Status | Typical use |
|---|---|---|---|---|---|---|
| Nova 2 Lite | Text, images, video, documents | Text | Fast, cost‑efficient reasoning; code interpreter and web grounding | Up to 1M tokens | GA in Bedrock | High‑volume chat, document Q&A, process automation |
| Nova 2 Pro | Text, images, video, documents | Text | Highest intelligence for multi‑step agentic tasks | Up to 1M tokens | Preview (via Nova Forge early access) | Complex analysis, software migrations |
| Nova 2 Omni | Text, images, video, speech | Text + images | Multimodal reasoning plus image generation/editing; 200+ text languages; speech understanding | Up to 1M tokens | Preview | Rich content creation, video analysis, voice + vision workflows |
| Nova 2 Sonic | Speech (plus cross‑modal voice↔text) | Speech | Real‑time, low‑latency conversational AI with multilingual “polyglot” voices | Up to 1M tokens | GA (regions vary) | Contact centers, live voice agents |
Sources: AWS What’s New posts and Nova docs.
Nova Forge: from fine‑tuning to “open training”
Nova Forge is a new path for enterprises that want expert, domain‑tuned models without starting from scratch. Builders can begin from pre‑, mid‑, or post‑training Nova checkpoints, blend proprietary data with Amazon‑curated corpora, run reinforcement fine‑tuning with their own reward functions, and apply responsible‑AI guardrails—all integrated with SageMaker AI. Early access to new Nova models (including Nova 2 Pro and Omni) is part of the package; GA launches first in US East (N. Virginia), with more regions to follow.
Industry reception points to a shift in priorities. As Amazon’s Rohit Prasad put it, today’s public leaderboards are “noisy” and less meaningful than real‑world utility—a stance consistent with Forge’s focus on bespoke capability. Reddit has already used Forge to train a platform‑expert safety model from decades of moderation data.
NVIDIA NVLink Fusion + AWS: the hardware story behind the software
NVIDIA’s NVLink Fusion is a chiplet and rack‑scale platform that lets custom ASICs snap into the NVLink scale‑up fabric and NVLink Switch—now in its sixth generation—alongside NVIDIA’s MGX rack architecture. For hyperscalers, it offers native memory semantics, in‑network reductions (SHARP), and a huge all‑to‑all topology—up to 72 custom ASICs at roughly 3.6 TB/s per‑ASIC, for 260 TB/s of scale‑up bandwidth. AWS says it will use NVLink Fusion with next‑gen Trainium4, and even connect Graviton CPUs and the Nitro System, to accelerate time‑to‑market for AI “factories.”
Why care? Because software velocity depends on infrastructure. The tighter scale‑up links inside a node (NVLink) and the standardization of rack‑level components (MGX) can reduce the time it takes AWS to stand up new, larger clusters—good news if you plan to train or serve Nova‑based agents at scale.
What this means for builders and ops teams
- If you’re building agents today, start small on Nova 2 Lite, then evaluate Pro for multi‑step tasks and Omni when you need image generation or video reasoning. Nova Sonic is the obvious pick for live voice.
- If you need a domain‑expert model (legal, safety, trading, claims): pilot Nova Forge using a narrow, high‑signal dataset; define reward functions and quality gates before you scale.
- Align infra plans with AWS’s roadmap. Trainium3 UltraServers are available now and target 4.4x more compute and 4x better energy efficiency versus the prior gen; NVLink Fusion will underpin the next Trainium era. Consider latency, memory, and parallelism choices early.
Competitive context—without the leaderboard drama
Coverage from WIRED and others notes Amazon’s claim that Nova 2 can match or exceed top‑tier models on select benchmarks—but the more interesting bet is that customization (via Forge) beats leaderboard climbing for enterprise outcomes. As The Verge summarizes, Amazon wants us to judge models by “real‑world utility,” not just eval sheets.
Sources
- AWS What’s New: Announcing Amazon Nova 2 foundation models in Amazon Bedrock
- AWS What’s New: Introducing Amazon Nova 2 Omni (Preview)
- AWS What’s New: Announcing Amazon Nova 2 Sonic for real‑time conversational AI
- AWS What’s New: Amazon Nova Forge: Build your own frontier models
- AWS Docs: What’s new in Amazon Nova 2 and What is Amazon Nova 2?
- About Amazon (re:Invent 2025): Live updates on new AI innovations
- AWS What’s New: Amazon EC2 Trn3 UltraServers
- NVIDIA blog: NVIDIA and AWS expand full‑stack partnership
- NVIDIA developer blog: AWS integrates AI infrastructure with NVIDIA NVLink Fusion
- Reporting and analysis: WIRED, The Verge, Reuters
