Browser-native local AI — which private/offline AI workflows should WorldScript optimize first?

[qnbs]

Why this exists

Issues #483 and #556 cover browser-native local inference and the lifecycle, integrity, performance, and Eco-Mode requirements for local Whisper/Kokoro workflows. This Discussion explores which private/offline AI workflows should be optimized first.

Current state

WorldScript Studio supports a mixed AI strategy: browser-local WebGPU/WASM, Ollama or other local-network services, and optional cloud providers. The product must not hide a local-to-cloud fallback. Model footprint, device capability, integrity, loading, cancellation, storage, and user-visible mode selection all matter, especially on low-end hardware.

Product decision space

Please compare smaller task-specific models with large generative models and discuss embeddings/RAG, grammar and style assistance, entity extraction, summarization, speech-to-text, text-to-speech, and Eco Mode. Consider WebGPU/WASM, Ollama/local-network workflows, cloud assistance, model download size, memory pressure, offline behavior, and graceful capability discovery.

Role perspectives

Role-perspective note: The viewpoints below are maintainer-curated, AI-assisted design lenses. They are not separate community members, votes, or evidence of consensus.

• ✍️ Privacy-focused author: local means visibly local, with no hidden provider fallback or unexplained data movement.
• ⚡ Low-end hardware: startup cost, RAM, thermal load, model size, cancellation, and usable degraded modes matter.
• 🌐 WebGPU/browser: capability detection, browser variance, worker/WASM behavior, and progressive enhancement matter.
• 🔐 Privacy: provider choice, telemetry, local-network trust, model integrity, and data residue must be explicit.
• 🧭 Product/model-footprint: prioritize high-value workflows with a sustainable download and maintenance budget.

Questions for the community

• Which local tasks should be optimized before general chat or large generation?
• Where do smaller specialized models outperform one large model in the writing workflow?
• What model size, startup time, and memory budget is acceptable on low-end devices?
• How should users understand WebGPU, WASM, Ollama, cloud, and Eco Mode choices?
• Which offline and privacy guarantees must be visible before execution?

Relationship to implementation

This Discussion links product exploration for #483 and #556. Those Issues remain the implementation and acceptance authorities; the Discussion does not change their priority, roadmap admission, acceptance criteria, or execution sequencing.