Models

Foundation models, accessed through Amazon Bedrock

CineCLI does not train or host its own foundation models. Instead, it routes each task to a suitable model from Amazon Bedrock — Anthropic Claude for language and reasoning, the Amazon Nova family for understanding, image and video generation, Amazon Titan for embeddings, and Stability AI for imagery. Every call runs under Amazon Bedrock Guardrails and our own governance controls.

Built on Amazon Bedrock Guardrails on every call Region-pinned
Model portfolio

Which models power which tasks

CineCLI selects a model per task rather than relying on a single model for everything. The table below summarizes how each model is used today.

Model Provider Used in CineCLI for Notes
Claude (Opus, Sonnet, Haiku family) Anthropic Script writing, story and scene development, reasoning, and content classification that supports moderation. Larger models for complex writing; smaller, faster models for classification and routing.
Amazon Nova Pro / Lite / Micro Amazon Text and multimodal understanding — analyzing scripts and reference media, structuring scenes, and assisting moderation. Tiered for cost and latency: Micro for fast text tasks, Pro for richer multimodal understanding.
Amazon Nova Canvas Amazon Image generation for storyboards, key frames and scene visuals. Image-generation model with built-in safety features such as watermarking.
Amazon Nova Reel Amazon Short video generation for scenes and motion shots. Video-generation model; outputs pass output screening and receive provenance credentials.
Amazon Titan (Embeddings, Image) Amazon Text and multimodal embeddings for search, continuity and retrieval; image generation where applicable. Embeddings power continuity look-ups and asset search across a project.
Stability AI models Stability AI Alternative image generation for storyboards and scene imagery. Used where its style or controls suit a shot; subject to the same guardrails and provenance.
Availability varies by AWS Region. The specific models and versions offered through Amazon Bedrock differ by Region, and CineCLI selects the appropriate model per task and per Region. If a model is not available in a customer's pinned Region, CineCLI routes the task to a suitable alternative or surfaces the limitation rather than silently substituting.
Why Amazon Bedrock

A managed, governable path to many models

Building on Amazon Bedrock lets CineCLI reach leading models through one managed interface, while keeping data handling, security and content controls consistent across providers.

Managed & in-region

Models run as a managed service in the AWS Regions you choose, so inference stays within your selected geography rather than crossing arbitrary boundaries.

Encrypted & private

Data is encrypted in transit and at rest. As described by AWS, content sent to Amazon Bedrock is not used to train the underlying base models.

Amazon Bedrock Guardrails

Configurable content filters apply policies for harmful content and denied topics, so the same safety rules can wrap every model regardless of provider.

Model evaluation

Bedrock's evaluation tooling helps us compare models on quality and safety for a given task before promoting one into the pipeline.

Provisioned throughput

For production workloads we can reserve dedicated capacity, giving predictable latency and throughput during peak production runs.

One interface, many models

A single managed API spanning Anthropic, Amazon, Stability AI and others lets us choose the best model per task without bespoke integrations.

Responsible model use

Every model call runs inside our safety envelope

No model is called directly. Each request is wrapped in Amazon Bedrock Guardrails and our own policy checks, and every generated asset is screened again before it can be used. Media that models produce carries provenance signals so it can be recognized as AI-generated.

  • Every model call wrapped in Amazon Bedrock Guardrails
  • Prompt and output screening on both ends of each call
  • No real-person likeness or voice without a verified consent record
  • Depictions of minors and other disallowed content are refused
  • Watermarking and C2PA content credentials on generated media
Responsible AI policy Acceptable Use Policy

Guardrails on every request

Input
Prompt and reference media screened before the model is called
Output
Generated text, image, video and audio screened before use
Likeness
Consent required for any real person's face or voice
Provenance
Watermark + C2PA credentials on generated media
Refusals
Disallowed requests are blocked, not best-effort fulfilled
Model governance

Versioned, evaluated and region-pinned

We treat model selection as a governed decision, with records of which model and version produced each result.

Versioning
Model and version recorded per generation for traceability
Evaluation
Models assessed for quality and safety before promotion
Region pinning
Inference constrained to your selected AWS Region
No base-model training
Customer data is not used to train base models
Built on, not endorsed by. CineCLI is built on and accesses models through Amazon Bedrock. References to Amazon, Anthropic, Stability AI and their models describe the technology CineCLI uses and do not imply any partnership, sponsorship or endorsement by those companies. Model names and availability are subject to change as Amazon Bedrock evolves.

Questions about models, data handling or Regions?

Talk to us about model selection, in-region deployment and how CineCLI governs every generation, or request access to try it.

Contact us Responsible AI