Nova (Amazon)

Amazon Nova is a family of foundation models (FM) and large language models (LLM) developed by the Amazon Artificial General Intelligence (AGI) organization and available through the fully managed Amazon Bedrock service. The family covers models for text understanding and generation, image, video, and document processing, as well as speech synthesis and recognition. Amazon Nova is positioned as the successor to the previous generation of Amazon Titan models and is integrated into the Amazon Web Services (AWS) cloud ecosystem.

History and development timeline

Before the launch of Nova, the Amazon Bedrock platform provided access to third-party models: Anthropic Claude, Meta Llama, Mistral, and others. Amazon Nova became AWS's new generation of in-house foundation models, succeeding the Amazon Titan line and designed from the start as a broad, tiered family for commercial use in the cloud infrastructure.

First generation (2024–2025)

The first generation of the Amazon Nova family was officially introduced on December 3, 2024, at the AWS re:Invent 2024 conference. The initial release included three understanding models — Nova Micro (text-only) and the multimodal Nova Lite and Nova Pro — as well as the generative models Nova Canvas (image generation) and Nova Reel (video generation).

In February 2025, cross-region inference profiles for Nova Micro, Lite, and Pro were extended to Europe and Asia Pacific. In March 2025, Amazon launched the nova.amazon.com website for direct experimentation with Nova models and released Nova Act — an agent and SDK for browser automation — as a research preview.

On April 30, 2025, the lineup gained the flagship Nova Premier model — the most capable model of the family, with a context window of 1 million tokens, intended for tasks with complex reasoning and for model distillation (transferring knowledge from a large model to a smaller one). Earlier that month, Nova Sonic was introduced — a speech-to-speech model supporting real-time dialogue. In October 2025, the family was joined by Nova Multimodal Embeddings — a unified embedding model for text, documents, images, video, and audio.

Second generation — Nova 2 (2025–2026)

At the AWS re:Invent 2025 conference in December 2025, the second generation — Amazon Nova 2 — was announced. The lineup includes the updated Nova 2 Lite (generally available) and Nova 2 Pro (preview) models with dynamic reasoning and extended context processing, the natively multimodal Nova 2 Omni (preview; simultaneous processing of text, images, video, and audio with text and image generation), and Nova 2 Sonic — the next-generation speech model. At the same time, Amazon introduced the Nova Forge service (an environment for custom model training) and announced the general availability of Nova Act for agentic browser workflows. The official Amazon Nova 2 technical report was published on December 2, 2025.

As of mid-2026, Nova 2 Lite and Nova 2 Sonic are generally available, while Nova 2 Pro and Nova 2 Omni remain in preview with early access provided to Nova Forge customers.

Summary timeline

Date	Event
December 2024	Amazon Nova announced at AWS re:Invent 2024: Nova Micro, Lite, Pro, Canvas, Reel
February 2025	Cross-region inference for Micro, Lite, and Pro extended to Europe and Asia Pacific
March 2025	Launch of nova.amazon.com; Nova Act (research preview)
April 2025	Release of Nova Premier (April 30; 1M-token context) and Nova Sonic (speech-to-speech); Nova Reel 1.1 (multi-shot videos up to 2 minutes)
June 2025	First-generation technical report published on arXiv (arXiv:2506.12103)
October 2025	Nova Multimodal Embeddings
December 2025	Nova 2 announced at AWS re:Invent 2025: Nova 2 Lite (GA), 2 Pro and 2 Omni (preview), 2 Sonic; Nova Forge; Nova Act becomes generally available; Nova 2 technical report (December 2)

Theoretical foundations and architecture

Base architecture of the understanding models

According to the technical report arXiv:2506.12103, the understanding models (Nova Micro, Lite, Pro, Premier) are based on the Transformer architecture described by Vaswani et al. The core multi-head self-attention mechanism is described by the formula:

$\mathrm{Attention}(Q,K,V)=\mathrm{softmax}\left(\frac{Q{K}^{\mathrm{\top }}}{\sqrt{d_k}}\right)V$

where $Q$, $K$, $V$ are the matrices of queries, keys, and values respectively, and $d_k$ is the key dimension.

The exact architecture parameters (number of layers, hidden-state size, number of attention heads, total parameter count) had not been disclosed in publicly available sources as of March 2026. This approach is typical of closed commercial models.

Multimodal processing in Nova Lite and Nova Pro is implemented by merging text, visual, and video tokens into a single sequence fed to a unified language model. Vision encoders convert images and video frames into sequences of tokens compatible with the text representation.

Architecture of the generative models

The generative models Nova Canvas (images) and Nova Reel (video) use the Latent Diffusion Model (LDM) architecture combined with Variational AutoEncoders (VAE) for computation in latent space. The diffusion process is described by the forward-noising equation:

$x_t=\sqrt{{\overline{\alpha }}_t}{x}_0+\sqrt{1-{\overline{\alpha }}_t}ϵ,ϵ\sim {𝒩}(0,I)$

where $x_0$ is the original image in latent space, $x_t$ is its noised version at step $t$, ${\overline{\alpha }}_t$ is the cumulative noise-schedule parameter, and $ϵ$ is standard Gaussian noise. A text encoder provides conditioning — grounding the generation on the text prompt.

Architecture of the speech models

Nova Sonic uses an architecture distinct from the text models: it combines a specialized speech encoder, a speech renderer, and the core multimodal language model into a single end-to-end pipeline. This allows the model to process speech input and generate speech output without intermediate conversion to text (although a text representation is used internally to ensure quality).

Training infrastructure

The Nova models were trained on distributed AWS Elastic Kubernetes Service (EKS) clusters using Amazon's own Trainium AI accelerators (TRN1 instances) as well as NVIDIA A100 and H100 GPUs.

To reduce the computational cost of training, specialized optimization methods were developed and applied:

• Super-Selective Activation Checkpointing (SSC) — an activation-saving method that, according to the technical report, reduces GPU memory consumption by approximately 50% with minimal recomputation overhead.
• In-CPU-Memory Checkpointing — caching intermediate weights (checkpoints) in CPU RAM before asynchronously offloading them to Amazon S3 cloud storage. According to Amazon, this approach reduced model state-saving time to 0.1 seconds on TRN1 instances.

Training pipeline and alignment

The training process of the Nova models consists of several stages typical of modern LLMs:

Pre-training

Large-scale training on a broad multilingual and multimodal data corpus covering more than 200 languages. The exact composition of the training data (volume, language ratios, specific sources) is not provided in public materials.

Supervised fine-tuning (SFT)

Fine-tuning on labeled instruction–response pairs, teaching the model to follow user instructions.

Alignment with reinforcement learning

To align model behavior with human preferences, two main algorithms are used:

Proximal Policy Optimization (PPO) — a reinforcement-learning-from-human-feedback (RLHF) algorithm that uses a separate reward model.

Direct Preference Optimization (DPO) — an alternative alignment method that does not require an explicit reward model. The DPO objective has the form:

${{ℒ}}_{\mathrm{D}\mathrm{P}\mathrm{O}}({\pi }_{\theta };{\pi }_{\mathrm{r}\mathrm{e}\mathrm{f}})=-{\mathbb{𝔼}}_{(x,y_w,y_l)\sim {𝒟}}[\log \sigma (\beta \log \frac{{\pi }_{\theta }(y_w\mid x)}{{\pi }_{\mathrm{r}\mathrm{e}\mathrm{f}}(y_w\mid x)}-\beta \log \frac{{\pi }_{\theta }(y_l\mid x)}{{\pi }_{\mathrm{r}\mathrm{e}\mathrm{f}}(y_l\mid x)})]$

where:

• ${\pi }_{\theta }$ is the parameterized policy (the model being trained);
• ${\pi }_{\mathrm{r}\mathrm{e}\mathrm{f}}$ is the frozen reference model;
• $x$ is the input prompt (context);
• $y_w$ and $y_l$ are the preferred (winner) and rejected (loser) responses, respectively;
• $\sigma$ is the logistic (sigmoid) function;
• $\beta$ is a hyperparameter controlling the strength of the penalty for deviating from the reference model (analogous to a KL-divergence penalty).

Model family

The model family is divided into tiers depending on the balance between performance, cost, and latency.

First-generation understanding models

Parameter	Nova Micro	Nova Lite	Nova Pro	Nova Premier
Input modalities	Text	Text, image, video	Text, image, video	Text, image, video
Output modality	Text	Text	Text	Text
Context window	128K tokens	300K tokens	300K tokens	1M tokens
Max output tokens	10K	10K	10K	10K
Language support	200+	200+	200+	200+
Fine-tuning	Yes	Yes	Yes	No
Release date	December 2024	December 2024	December 2024	April 2025
Primary purpose	Minimal latency and cost for text tasks	Basic multimodal analysis	Optimal balance for complex reasoning and agentic tasks	Maximum accuracy; teacher model for distillation

Source: AWS AI Service Cards; arXiv:2506.12103.

Optimized languages (15): English, German, Spanish, French, Italian, Japanese, Korean, Arabic, Simplified Chinese, Russian, Hindi, Portuguese, Dutch, and Turkish, plus Hebrew.

Nova Premier is designed for tasks requiring deep context understanding, multi-step planning, and work with large volumes of data: codebases, documents of more than 400 pages, and videos up to 90 minutes long.

Second-generation models (Nova 2)

Nova 2 Lite and Nova 2 Pro were released in December 2025. Both support extended thinking — a mechanism in which the model performs multi-step reasoning before generating a response. The reasoning depth is controlled by a reasoning-effort setting with low, medium, and high levels. The context window is extended to 1 million tokens. Native tools are built in: web search with grounding and a code interpreter.

Nova 2 Omni is a natively multimodal model that can simultaneously accept text, images, video, and audio as input and generate text and images. Its context window is 1 million tokens; it is available in preview.

Nova 2 Sonic is the next-generation speech model. It supports seven languages — English, French, Spanish, German, Italian, Portuguese, and Hindi — including "polyglot" voices that can switch languages within a single conversation, and a context window of up to 1 million tokens. It introduces asynchronous tool calling — the model continues processing new input while external tools run in the background — and cross-modal switching between voice and text within a session.

Parameter	Nova 2 Lite	Nova 2 Pro	Nova 2 Omni	Nova 2 Sonic
Input modalities	Text, image, video	Text, image, video	Text, image, video, audio	Audio (speech)
Output modality	Text	Text	Text, image	Audio (speech)
Context window	1M tokens	1M tokens	1M tokens	Up to 1M tokens
Extended thinking	Yes	Yes	Yes	—
Native tools	Web grounding, code interpreter	Web grounding, code interpreter	—	Asynchronous tool calling
Status (mid-2026)	Generally available	Preview	Preview	Generally available
Release date	December 2025	December 2025	December 2025	December 2025

Source: AWS Blog; Amazon Science; AWS product page.

Generative models

Nova Canvas — an image-generation model. It supports text and image input (PNG, JPEG). Output resolution — up to 4.19 million pixels (e.g., 2048×2048 or 2816×1536 pixels). Maximum prompt length — 1,024 characters. Inpainting (replacing a region of an image) and outpainting (extending an image beyond its borders) are supported.

Nova Reel — a video-generation model. Output resolution: 1280×720 at 24 frames per second. At launch, generated videos were up to 6 seconds long; Nova Reel 1.1 (April 2025) extended this to multi-shot videos of up to 2 minutes. Camera control is supported. Access is via an asynchronous API (Asynchronous Invoke Model API).

Speech models

Nova Sonic (April 2025) — a speech model with a bidirectional streaming API. It supports function calling and grounding on corporate data via Retrieval-Augmented Generation (RAG). Audio watermarking is enabled by default. The maximum connection duration is 8 minutes, with the option to resume; the default limit is 20 concurrent connections per client.

Nova 2 Sonic (December 2025) — the next-generation speech model with improved recognition of short utterances, telephone audio (8 kHz), and accents, and with integrations into Amazon Connect, telephony providers (Vonage, Twilio, AudioCodes), and conversational frameworks (LiveKit, Pipecat).

Evaluation and benchmarks

Nova models were evaluated using automated benchmarks, including the LLM-as-a-judge approach (using a language model as the evaluator). According to the HKU SPACE AI Hub study, the Arena-Hard-Auto metric shows a 98.6% correlation with expert judgments.

First-generation benchmarks

Self-reported figures from the technical report arXiv:2506.12103:

Benchmark	Nova Micro	Nova Lite	Nova Pro
MMLU (5-shot)	77.6	80.5	85.9
MATH	69.3	73.3	76.6
IFEval	87.2	87.5	92.1

Source: arXiv:2506.12103 (developer-reported results).

An independent MT-Bench comparison from the AWS ML Blog / HKU SPACE AI Hub study (scores as reported):

Benchmark	Nova Pro	Nova Lite	Claude 3.5 Sonnet (Oct 2024)	GPT-4o (Nov 2024)
MT-Bench (text)	79.7	77.7	76.7	77.5
MT-Bench (multimodal)	63.7	60.7	61.6	55.0

The results show a performance hierarchy within the family (Premier > Pro > Lite > Micro). The gap is most pronounced in the Math and Reasoning categories; in Roleplay and Extraction tasks, the smaller models perform close to the larger ones.

Second-generation benchmarks (Nova 2)

Figures from the Amazon Nova 2 technical report (conditions: internal measurements, 0-shot / CoT where indicated):

Benchmark	Nova 2 Lite	Nova 2 Pro	Claude Haiku 4.5	GPT-5 Mini	Gemini 2.5 Flash
MMLU-Pro (acc)	80.9	81.6	80.0	83.7	83.2
GPQA-Diamond (acc)	79.6	81.4	73.0	82.3	82.8
AIME 2025 (acc)	91.0	92.3	80.7	91.1	72.0
LongCodeBench 1M (acc)	84.0	84.0	—	—	78.0

Source: Amazon Nova 2 Technical Report, Table 1.

Agentic benchmarks (Nova 2 Pro): SWE-Bench Verified — 70.0%; τ²-bench Verified Telecom — 92.7%.

Multimodal benchmarks (Nova 2 Omni): VideoMME — 77.9%; MMMU Pro — 61.4%.

In customer-support task evaluations, Nova 2 Lite scored 9.63 ± 0.27 on a ten-point scale in the Problem Identification metric, significantly ahead of the first-generation Nova Lite (8.57 ± 0.46).

Note. When comparing results against competing models, one must account for differences in prompting conditions, model versions, and measurement dates. The first- and second-generation benchmarks are Amazon self-reports; independent verifications (FloTorch, Artificial Analysis) broadly confirm the claimed price/performance ratio but record gaps behind leading competitors on some accuracy metrics.

Inference speed

According to Amazon's internal measurements (via the Bedrock API):

Model	Inference speed (tokens/s)
Nova Micro	≈210
Nova Lite	≈157
Nova Pro	≈100
Nova 2 Omni	>200

Pricing

All models are available through Amazon Bedrock on a pay-per-token basis (first-generation list prices as of March 2026):

Model	Input tokens (USD / 1M)	Output tokens (USD / 1M)
Nova Micro	$0.035	$0.14
Nova Lite	$0.06	$0.24
Nova Pro	$0.80	$3.20
Nova Premier	$2.50	$12.50

For comparison: Anthropic Claude 3.5 Sonnet was priced at $3.00 / 1M input and $15.00 / 1M output tokens at the time of the Nova release. Nova 2 models are priced separately; current rates are published in the Amazon Bedrock pricing documentation.

Customization and fine-tuning

The AWS infrastructure provides several methods for adapting the base Nova models to specialized domains. In the second generation, the primary tool for this is the Amazon Nova Forge environment; custom models built with it are branded "Novellas," and training can use customer-defined reinforcement-learning "gyms" — synthetic environments reflecting real workflows.

Continued Pre-Training (CPT) and Mid-Training

Nova Forge opens access to intermediate training checkpoints of the models (e.g., pretraining-text-RD and pretraining-text-CE). This makes it possible to continue self-supervised learning on corporate data corpora, with careful learning-rate selection to prevent catastrophic forgetting.

Parameter-Efficient Fine-Tuning (PEFT)

Updating only a small subset of model weights through low-rank adapters — Low-Rank Adaptation (LoRA). This approach substantially reduces compute requirements compared with full fine-tuning. Multimodal fine-tuning is supported for Nova Lite and Pro.

Reinforcement Fine-Tuning (RFT)

Custom models can be further trained with reinforcement methods based on industry-specific reward metrics, including using another LLM as a judge (LLM-as-a-judge).

Model Distillation

The Model Distillation feature allows Nova Premier or Nova Pro to be used as a teacher model for training smaller student models — Nova Lite and Nova Micro. This reduces inference compute costs while retaining a substantial share of the large model's quality.

Applications and integration

Nova models are integrated into Amazon compute services (AWS Step Functions, AWS Lambda, Amazon Q Developer); in the US, the models can also be tried directly at nova.amazon.com. The main documented application scenarios:

Agentic workflows

Multi-step task execution using Bedrock Agents and external tool invocation (function calling). Using the ReAct (Reasoning and Acting) architectural pattern together with frameworks such as LangGraph, Nova models coordinate database analytics, form SQL queries from natural language, and manage multi-part processes. Nova Act — generally available since December 2025 and built on a custom Nova 2 Lite model — provides browser UI-workflow automation with a claimed reliability of over 90%.

Retrieval-Augmented Generation (RAG)

Integration with Bedrock Knowledge Bases for grounding responses on corporate data. Nova 2 models support native web grounding as a built-in tool.

Document processing

Supported input document formats: PDF, CSV, DOC, DOCX, XLS, XLSX, HTML, TXT, MD (except Nova Micro, which accepts only text input). Nova Premier can process videos up to 90 minutes long within a single request.

Code generation and debugging

Supported programming languages: Python, Java, JavaScript, C#, TypeScript, SQL.

Speech interfaces

Nova Sonic and Nova 2 Sonic are used to build real-time voice agents integrated with Amazon Connect.

Model evaluation (LLM-as-a-judge)

Nova is used as a judge model when evaluating the outputs of other generative models on the Amazon SageMaker platform.

Limitations and open problems

• Closed architecture. Amazon does not disclose the parameter counts, detailed architectural decisions, or training-data composition, which substantially limits independent reproducibility. Nova model weights are not available for download or deployment outside the AWS infrastructure (on-premise deployment is impossible).
• Output limit. The maximum number of output tokens for all first-generation understanding models is capped at 10K, which may be insufficient for long-document generation tasks.
• RFT limitations. Reinforcement Fine-Tuning does not support multi-turn dialogues or multimodal data; the recommended share of positive examples in the dataset is at least 5%.
• Nova Sonic limitations. Maximum connection duration — 8 minutes; a default maximum of 20 concurrent connections per client.
• Regional availability. Nova Premier is available only in US regions (US East N. Virginia, US East Ohio, US West Oregon) via cross-region inference; Micro, Lite, and Pro have been available in Europe and Asia Pacific via cross-region inference since February 2025; the Nova 2 models launched with a limited list of deployment regions.
• Metric sensitivity. Scores on synthetic benchmarks do not always correlate with quality in production settings, where strict adherence to corporate policies and the absence of hallucinations in multi-step outputs are critical.
• Hallucinations. The models exhibit limitations typical of LLMs: factual errors in generated responses are possible. To reduce risks, the use of Bedrock Guardrails and RAG is recommended.
• Comparative objectivity of benchmarks. Amazon presents comparisons with competing models based on data published by those models' own developers, which does not always provide a single controlled experimental baseline.

Ethical and regulatory aspects

AWS declares adherence to Responsible AI principles in developing the Nova models. Specific safety measures include:

• Built-in content moderation.
• Watermarking of Nova Sonic audio outputs.
• Evaluation across risk categories: safety, privacy, veracity, transparency, as well as CBRN (chemical, biological, radiological, and nuclear) weapons proliferation and offensive cyber operations.
• Integration with Amazon Bedrock Guardrails for filtering inputs and outputs.
• Evaluation of the Nova Premier model against the Amazon Frontier Model Safety Framework.
• Red teaming — internal and external adversarial-robustness testing (307 techniques, according to the technical report).
• Content-moderation F1 score of 85.84 on the Aegis benchmark (per the technical report).

AI Service Cards for Nova Micro, Lite, Pro, Premier, and Sonic are published on docs.aws.amazon.com as responsible-use documentation.

Outlook and research directions

The Nova 2 family marks a transition to models with extended thinking/reasoning capabilities, conceptually comparable to Chain-of-Thought (CoT) and similar mechanisms in other model families. The built-in web search and code interpreter as native tools (rather than third-party plugins) represent an architectural shift toward agentic systems.

Directions for further development outlined in Amazon's public materials:

• Expansion of multimodality (the Omni model as a unified "everything-to-everything" architecture).
• Hybrid reasoning with adaptive depth depending on task complexity.
• Open training on customer data (Nova Forge) at all stages of pre-training.
• Distillation for edge devices.
• Expansion of the safety toolkit.

The themes of the Amazon Nova AI Challenge, which AWS runs among university teams, include automated adversarial testing, safety alignment, and multi-turn jailbreaks — evidence of investment in the reliability of the model family.

External links

• Official Amazon Nova documentation
• AI Service Cards for Nova Micro, Lite, Pro, and Premier
• AI Service Card for Nova Sonic
• Amazon Nova launch announcement (December 2024)
• Amazon Nova 2 launch announcement (December 2025)

See also

• Transformer (architecture)
• Reinforcement Learning from Human Feedback (RLHF)
• Retrieval-Augmented Generation (RAG)
• Claude

Literature

• Amazon Artificial General Intelligence. The Amazon Nova Family of Models: Technical Report and Model Card. arXiv:2506.12103v1, 2025. https://arxiv.org/abs/2506.12103
• Amazon Artificial General Intelligence. Amazon Nova 2: Multimodal Reasoning and Generation Models — Technical Report and Model Card. Amazon Science, December 2, 2025. https://www.amazon.science/publications/amazon-nova-2-multimodal-reasoning-and-generation-models
• Vaswani, A. et al. (2017). Attention Is All You Need. NeurIPS. https://arxiv.org/abs/1706.03762
• Rafailov, R. et al. (2023). Direct Preference Optimization: Your Language Model is Secretly a Reward Model. NeurIPS. https://arxiv.org/abs/2305.18290
• Ouyang, L. et al. (2022). Training language models to follow instructions with human feedback (InstructGPT). NeurIPS. https://arxiv.org/abs/2203.02155
• Rombach, R. et al. (2022). High-Resolution Image Synthesis with Latent Diffusion Models. CVPR. https://arxiv.org/abs/2112.10752
• Hu, E.J. et al. (2021). LoRA: Low-Rank Adaptation of Large Language Models. https://arxiv.org/abs/2106.09685