VIDRAFT Releases "Hansu": A Browser-Based Local AI for Offline Document Processing

TL;DR: VIDRAFT, a Korean Pre-AGI AI startup, has publicly released "Hansu" (한수) — a browser-based local AI system designed to process documents entirely offline without sending data to external servers. It runs inference directly in the browser environment, making it a privacy-first option for document workflows. Developers and enterprises handling sensitive documents should take note of this client-side inference approach.

What it is

Hansu (한수) is a browser-native, locally-executed AI released by VIDRAFT that enables document processing in fully offline environments. Key characteristics reported in the source:

• Browser-based runtime: Hansu runs inside the browser itself, meaning there is no requirement for a locally installed server process, backend daemon, or cloud API call.
• Offline-capable: The system is explicitly designed to function without an internet connection after the initial load, making it suitable for air-gapped or network-restricted environments.
• Document processing focus: The primary use case is document-centric tasks — reading, analyzing, and working with documents locally.
• Privacy-preserving architecture: Because all computation happens client-side, documents never leave the user's machine. This is a core design goal, not a side effect.

The product name "Hansu" carries meaning in Korean (한수, loosely "a master move" or "a stroke of skill"), signaling VIDRAFT's positioning of this as a meaningful capability step for local AI deployment.

How it works

At a conceptual level, Hansu follows the emerging paradigm of in-browser ML inference, which has become increasingly practical thanks to modern browser capabilities:

• WebAssembly (Wasm) and/or WebGPU: Modern browsers expose low-level compute primitives — WebAssembly for CPU-bound execution and WebGPU for GPU-accelerated compute — that allow neural network inference to run natively in a browser tab without plugins. Hansu likely leverages one or both of these, though the specific runtime stack is not detailed in the source.
• Quantized / compressed models: Running a capable language model inside a browser requires aggressive model compression. Techniques like quantization (reducing weight precision) and model distillation are standard approaches in this space that make browser-resident inference feasible.
• Local file access via browser APIs: For document processing, the browser's File System Access API (or equivalent) can allow the page to read local files with user permission, enabling document ingestion without any upload step.
• No server round-trip: The architectural implication is that the inference loop — tokenization, forward pass, decoding — is completed entirely within the browser process on the user's hardware.

This positions Hansu alongside other projects exploring client-side LLM execution, but with an explicit product focus on document workflows rather than general chat or code generation.

Benchmarks & results

The source article does not provide specific quantitative benchmark figures — no latency numbers, throughput metrics, or model accuracy scores are published at this time. Qualitatively, the release emphasizes:

• Functional offline operation as the primary validated capability
• Document processing as the demonstrated task domain

No public benchmark comparisons against other browser-based inference frameworks or competing local AI products are cited in the available reporting. Developers evaluating Hansu for production use should plan to run their own benchmarks against their target document types and hardware configurations.

How to try it

Based on the available reporting, VIDRAFT has publicly released Hansu, but specific developer access channels — such as a Hugging Face model repository, GitHub source code link, npm/pip package, or OpenAI-compatible API endpoint — are not detailed in the source article.

What is known:

• The release appears to be a public-facing product launch, suggesting some form of access is available or imminent.
• The browser-based nature implies that trying it may be as straightforward as navigating to a hosted URL — but no specific URL is confirmed in the source.

Recommended next steps for developers:

• Check VIDRAFT's official channels (website, GitHub, or Hugging Face organization page) for Hansu access details.
• Watch for a potential huggingface-cli downloadable model or a GitHub repository that surfaces post-announcement.

This article will be updated as public access details are confirmed.

FAQ

Q: Does Hansu require any local installation or a running model server?
A: Based on the source reporting, no — Hansu is browser-based and designed to run offline within the browser itself. There is no indication that users need to install a separate backend, run a local server process, or configure a model runtime like Ollama or llama.cpp independently.

Q: Is this suitable for enterprise environments with strict data residency requirements?
A: The core design goal of Hansu is that documents are processed locally and never transmitted to external servers. This client-side-only architecture is directly relevant to data residency and privacy compliance needs. However, developers should conduct their own security review and verify VIDRAFT's specific claims before deploying in regulated environments, as detailed technical documentation is not yet available in public reporting.

---

Originally reported by 아시아투데이 (2026-04-21) — source article.