The Amazon Nova Family of Models: Technical Report and Model Card (Amazon, 2024–2025)

The Amazon Nova Family of Models: Technical Report and Model Card is the document in which Amazon's Artificial General Intelligence organisation describes the first generation of its Amazon Nova foundation models (Source: https://arxiv.org/abs/2506.12103). The arXiv version (2506.12103v1) is dated March 17, 2025; Amazon first published the report on its science site alongside the December 2024 launch of the models on Amazon Bedrock. The report covers five models: Nova Pro, described as a multimodal model with "the best combination of accuracy, speed, and cost"; Nova Lite, a low-cost multimodal model; Nova Micro, a text-only low-latency model; Nova Canvas, an image-generation model; and Nova Reel, a video-generation model. It does not cover the later Nova Premier or Nova 2 models, nor the March 2025 Nova Act agent system.

Models

Section 1 describes Pro, Lite and Micro as offering "frontier intelligence" in their price tiers, with Micro presented as setting new standards in its tier on MMLU, GPQA, MATH and BIG-Bench Hard, and Pro and Lite accepting text, images, documents and video as input. Context lengths are 128k tokens for Micro and 300k for Lite and Pro. The report lists speed, agentic capability (function calling and multi-step tool use), long-context understanding and customisation (fine-tuning and distillation on Bedrock) as design goals. Canvas is described as generating "professional grade images with rich customization controls", and Reel as offering "high-quality outputs, customization, and motion control". Section 6 covers training infrastructure; the report gives no parameter counts.

Evaluation approach

Section 2 reports benchmarks for Pro, Lite and Micro and for "select publicly-available models", citing existing public results where available and measuring the comparison models itself where not. Results measured by Amazon are marked "M" in the tables and were run "after Amazon Nova models completed training" via the Bedrock API (Claude and Meta models), the OpenAI API or the Gemini API. Where a result is an average of binary scores, the report gives a 95% confidence interval from a Gaussian approximation. Shot settings and prompting differ by model group and are listed beneath each group in the tables; prompts and scoring rules are given in Appendix B.

The benchmarks are grouped as core text capability (MMLU, ARC-C, DROP, GPQA, MATH, GSM8K, IFEval, BIG-Bench Hard, and FLORES-200 translation), core multimodal capability (MMMU, ChartQA, DocVQA, TextVQA, VATEX, EgoSchema), agentic workflows (Berkeley Function Calling Leaderboard v3 as of its November 17, 2024 update; VisualWebBench, MM-Mind2Web and GroundUI-1K), long context (a text needle-in-a-haystack test shown as a heatmap, SQuALITY and LVBench), functional expertise (HumanEval, FinQA and CRAG under Amazon's own scoring), and runtime performance in tokens per second. Section 3 evaluates Canvas with the automated TIFA and ImageReward metrics and with human pairwise preference against DALL·E 3 and Imagen 3; Section 4 evaluates Reel by human pairwise preference against Runway Gen3 Alpha and Luma 1.6.

Findings

On the core text table, Nova Pro is reported at 85.9 on MMLU (0-shot CoT), 94.8 on ARC-C, 46.9 on GPQA, 76.6 on MATH, 94.8 on GSM8K, 92.1 on IFEval and 86.9 on BBH; the report presents these as competitive with Claude 3.5 Sonnet, Gemini 1.5 Pro and GPT-4o at higher speed, and Micro as leading its tier. On multimodal benchmarks Nova Pro is reported at 61.7 on MMMU and 89.2 on ChartQA, below Claude 3.5 Sonnet on MMMU (70.4) and ChartQA (90.8). On BFCL v3, Nova Pro's overall accuracy of 68.4 is reported beside GPT-4o (Aug) at 68.9 and Claude 3.5 Sonnet (Jun) at 61.3. The report states that Nova Pro and Lite "surpass" other models on LVBench and on the agentic multimodal benchmarks, where several comparison values were measured by Amazon. Human raters preferred Nova Canvas to DALL·E 3 for overall image quality in 54.5% of pairs and to Imagen 3 in 48.2%; Nova Reel won 56.4% of video-quality comparisons against Gen3 Alpha and 51.1% against Luma 1.6, and 67.0% and 74.7% for video consistency.

Responsible AI

Section 5 sets out eight Responsible AI dimensions, says Amazon signed the White House voluntary commitments, and describes pre-training data curation, SFT and RLHF alignment, and runtime input and output moderation models. It names the public benchmarks BOLD, RealToxicityPrompts and MM-SafetyBench and proprietary dynamically-updated benchmarks, but reports no scores. Red teaming combined internal exercises (more than 300 attack techniques), external firms including ActiveFence (more than 9,700 adversarial prompts across 20 categories) and other named vendors covering CBRN and cybersecurity, and automated red teaming; findings are described as fed back into development without quantitative results.

Reported results

Evaluation results the document states for which there is no separate benchmark page; results on benchmarks that have a page are recorded there and cite this page. Scores and settings are as printed in the document.

Provenance

The page summarises the arXiv HTML (v1, March 17, 2025) as captured on September 21, 2026. All Nova figures are Amazon's own; comparison figures are either the other developers' published numbers or Amazon's post-training measurements via public APIs, as marked in each table. Confidence is medium: a single developer-authored source with fast-decay benchmark claims.

Relationships

• supports: Amazon Nova
• instance-of: AI System Cards
• related: Amazon, Berkeley Function Calling Leaderboard (BFCL), MMLU, GPQA, HumanEval, Claude 3.5 Haiku, Llama 3 (Meta), GPT-4 Family (OpenAI)