Amazon Nova

Original Article ↗ Hacker News Discussion ↗

Amazon has introduced its Nova family of AI foundation models on AWS, aiming to compete with frontier systems like GPT-4 and Claude by emphasizing speed and low cost over absolute state-of-the-art quality. Commenters note that Nova Pro appears roughly competitive on some benchmarks and particularly strong on certain vision tasks, while Nova Lite and Micro target “good enough” use cases at a fraction of typical token prices. Reactions focus on trade-offs between quality and price, Bedrock’s complexity and lock-in, the absence of audio and EU-hosted options, and Amazon’s broader strategy of offering both its own “house brand” models and premium third‑party options like Anthropic’s Claude.

Model quality, benchmarks, and positioning

  • Technical report and benchmark comparisons show Nova aiming more at speed/price than absolute top scores.
  • Nova Pro appears SOTA-comparable on some tasks (e.g., big win on GroundUI-1K) but not dominant overall.
  • Several commenters stress that benchmarks are easy to game and don’t match performance on novel tasks.
  • On some public leaderboards, Nova Pro is roughly on par with mid-tier open models (e.g., Yi Coder 9B), which some find underwhelming.
  • Others note “good enough” lightweight models are increasingly attractive if they’re cheap and fast.

Pricing, value, and use cases

  • Nova models are substantially cheaper than Anthropic’s Claude line; one estimate: roughly ~1.5 orders of magnitude cheaper per token in some tiers.
  • Independent latency/quality sites report Nova as cheap and fast with slight quality drop vs top models.
  • Some see this as ideal for cost-sensitive workloads, “good enough” features, or when already on AWS.
  • Critics argue quality matters more than price; they see Nova as a “me too” or DOA product that will lag frontier models.

Relationship to Anthropic and broader AWS strategy

  • Many view Nova as an “Amazon Basics” house brand alongside premium third-party options like Anthropic in Bedrock.
  • Rationale given: customers want choice, AWS wants to keep AI spend on its cloud, and different models have different strengths and error profiles.
  • Bedrock is seen by some as the best way to access Claude 3.5 while also trying Nova.

APIs, UX, and setup complexity

  • Bedrock/Nova setup is described as cumbersome (many IAM/region/access steps) compared to simple API-key flows.
  • Debate over the value of Bedrock’s abstraction:
    • Pro: single platform, unified billing/SSO, multiple models via one API.
    • Con: confusing IAM model, unintuitive product descriptions, heavy stacks for OpenAI-compatible proxies.
  • Amazon’s UI/UX and marketing copy are widely criticized as clunky and jargon-heavy; some argue AWS optimizes for functionality and speed of feature delivery over polish.

Capabilities, limitations, and openness

  • Nova is multimodal for text and video but explicitly does not handle audio; some speculate this is a deliberate product separation from ASR.
  • Speech-to-speech and “any-to-any” models were mentioned as coming later.
  • No embedding endpoints announced; no parameter counts in the main announcement; param estimates shared via third-party links.
  • Models are closed-source; some wish for truly open counterparts like OLMo.

Compliance, hosting, and regional issues

  • Lack of clearly EU-hosted variants is seen as a missed opportunity given data residency and retention rules.
  • There is disagreement over whether EU hosting meaningfully mitigates US CLOUD Act concerns, but it clearly matters for many European customers’ compliance policies.