Zera & Local Models Policy
Model downloads, local execution, device requirements, data boundaries, licenses, and limitations of Zera and related local features.
1. Download and local runtime
Zera is distributed as a model download, including the 24B, 4-bit MLX distribution linked from the product. Model files require a compatible local runtime and sufficient device resources. A download page, a model name in a chooser, or an uploaded filename is not itself a running local inference engine.
Local features must identify a supported runtime and its connection or model-loading requirements. A browser cannot be assumed to execute any arbitrary model archive directly. Follow the product’s actual setup instructions and verify that the local runtime is responding before relying on local execution.
2. Device resources and compatibility
Storage needs, memory use, context limits, speed, and supported operating systems vary by model format, runtime, quantization, and hardware. Review the current model card and runtime instructions rather than assuming every phone or Mac can load the complete model.
Allow storage for downloads, unpacked files, caches, and temporary data. An interrupted archive may be incomplete even if a browser reports that the request started. Compare downloaded files with the published manifest or checksums when available. Avoid repeatedly downloading large files on metered connections.
3. What local means
When inference is performed by a local runtime on your device, the model request is processed there rather than by a hosted model for that inference. This statement describes the runtime path; it does not claim that the entire website, browser, operating system, or download service is offline.
Signing in, syncing chats, downloading models, opening sources, connected plugins, and cloud features can still contact external services. Local runtimes may have their own telemetry or update requests. Review their settings and verify the route used for each task. Do not assume a hosted fallback has the same privacy boundary.
4. Cloud features and saved content
Selecting a local model should not be interpreted as authorizing an undisclosed cloud substitution. If a local connection is unavailable, the product should disclose that state and request an appropriate choice rather than label a hosted answer as local.
Saving a conversation to a cloud account, using research or images, or asking a connected service to act may separately transmit content. Private mode and local inference are different controls. The Privacy Policy describes hosted processing and saved history; neither label overrides an external provider’s practices.
5. Model and component licenses
Model weights, tokenizers, runtime software, training data disclosures, and bundled components can have separate licenses. The authoritative license is the one distributed with the specific model or component. A free download does not automatically permit every commercial use, redistribution, or modification.
Keep required notices and follow license restrictions when distributing a derivative or combining components. This policy does not replace the model’s license or grant rights that Zexiry does not hold. Check the model card on the linked Hugging Face repository for current files and notices.
6. Output quality and responsible use
Local models can produce inaccurate, biased, insecure, or harmful output and may lack current information. Local execution does not make an answer inherently reliable. Apply the AI Accuracy & Reliance and Acceptable Use policies when using a local model with our products.
Do not use an unverified output for high-impact decisions or present a local model as a professional, human reviewer, or certified system. You remain responsible for prompts, final actions, and the content you choose to publish.
7. File safety and removal
Obtain model files from the published source. Treat executable scripts, runtime installers, and untrusted model serialization formats with caution. Do not upload unrelated personal files or credentials to a local-model setup flow. Where a flow accepts a file, inspect whether it is processed locally or transferred before proceeding.
Deleting a model from your device or clearing a runtime’s cache is separate from deleting saved chats or account data. Uninstalling a runtime may leave model files and logs behind. Follow the runtime’s removal instructions and keep backups of anything you intend to retain.
8. Zeneral and experimental releases
Zeneral is an ongoing architecture effort focused on lower-VRAM model training. Research direction is not a guarantee of a particular memory reduction, training speed, accuracy, supported workload, or production readiness. Evaluate published measurements and their conditions before making comparisons.
Experimental models and training code may change or be discontinued. Use data you are entitled to train on and comply with its license and privacy obligations. Contact mail@zexiry.com for questions about the supported release, model files, or research materials.
Related policies and contact
For related responsibilities, see our Acceptable Use, AI Accuracy & Reliance, Account & Security, and Subscription & Billing policies. Contact mail@zexiry.com for questions about this supplement.