VIDRAFT Launches "Hansu": A Browser-Based Local AI Built for Data Sovereignty

TL;DR: VIDRAFT, a Korean Pre-AGI AI startup, has publicly released "Hansu" (한수) — a browser-based local AI that runs entirely client-side without sending data to external servers. Designed around the principle of data sovereignty, it gives developers and organizations a way to deploy AI inference directly in the browser, keeping sensitive data on the user's device. If you care about privacy-preserving AI deployment or want to explore local LLM execution without a backend dependency, this is worth watching.

What it is

Hansu (한수) is a browser-native local AI service developed by VIDRAFT. Key characteristics drawn from the source:

• Runs in the browser, locally: Inference happens client-side, meaning the model executes on the end user's machine rather than on a remote server.
• No data leaves the device: The architecture is explicitly designed around data sovereignty — user inputs and outputs stay local, which directly addresses regulatory and enterprise privacy concerns.
• Framed as a "data sovereignty-type service": VIDRAFT is positioning Hansu not just as a product but as a category of deployment — one where the user, not the cloud provider, retains control over their data.

This puts Hansu in the same general space as projects like browser-based WASM/WebGPU LLM runtimes, but as a productized, named offering from a startup with a specific Pre-AGI research focus.

How it works

At a high conceptual level, browser-based local AI of this kind typically relies on one or more of these mechanisms — and Hansu appears to follow this general architectural pattern:

• Client-side model execution: Rather than making API calls to a remote inference endpoint, the model weights are loaded into the browser environment and run using available hardware acceleration (such as WebGPU or WebAssembly).
• No network round-trips for inference: Once the model is loaded, queries are processed locally. This eliminates the latency of server communication and, crucially, the exposure of query data to third-party infrastructure.
• Data stays on-device: Because there is no server-side component handling user input, the privacy boundary is the user's own machine. This is structurally different from "private cloud" deployments — it's genuinely local.

VIDRAFT's framing of this as a "data sovereignty" model suggests the design intent goes beyond a technical convenience feature — it's positioned as a trust architecture choice, particularly relevant for healthcare, legal, finance, and public-sector use cases where data residency requirements are strict.

Benchmarks & results

The source article (a headline-level report from 전자신문) does not include specific quantitative benchmarks — no latency figures, throughput numbers, model size disclosures, or accuracy comparisons were reported in the available coverage.

What can be said qualitatively:

• VIDRAFT chose to make this a public release, suggesting the team considers it ready for external evaluation.
• The emphasis on browser-based local execution implies a deliberate trade-off: the approach prioritizes privacy and deployability over raw performance headroom, which is a meaningful architectural stance in the local AI space.
• As public benchmark data becomes available (e.g., on Hugging Face model cards or official documentation), those numbers will be the right place to evaluate inference speed and model capability.

If you run your own benchmarks after access, sharing results in the community would be genuinely useful.

How to try it

Based on the available source reporting, Hansu has been publicly released, but specific developer access channels — such as a Hugging Face repository, GitHub link, npm package, or OpenAI-compatible API endpoint — were not detailed in the coverage available at time of writing.

What to do right now:

• Check VIDRAFT's official channels for a direct link to the Hansu browser interface or any accompanying SDK/repository.
• Watch for a Hugging Face organization page under VIDRAFT for model weights or documentation.
• The browser-based nature of the service suggests you may be able to try it directly via a web URL without any install step — look for that on their official site.

This article will be updated if public repository or API access details are confirmed.

FAQ

Q: Does "browser-based local AI" mean the model weights are downloaded to my machine every session?
A: In most browser-local AI implementations, weights are fetched from a CDN or server on first load and cached in the browser (via IndexedDB or similar mechanisms). Subsequent sessions use the cached weights. The key point is that inference itself runs locally — your input data never leaves the browser sandbox. VIDRAFT's specific caching behavior for Hansu has not been detailed in public reporting yet.

Q: What kinds of use cases is Hansu designed for?
A: VIDRAFT explicitly frames Hansu as a "data sovereignty-type service," which points toward enterprise and regulated-industry use cases — scenarios where sending data to a cloud LLM API is legally or contractually problematic. Think internal document analysis, sensitive customer interactions, or on-device personal assistants where data residency matters.

Q: Is Hansu open-source?
A: The source reporting does not confirm whether Hansu's underlying model or runtime code is open-source. Given VIDRAFT's Pre-AGI research positioning, watch their GitHub and Hugging Face presence for any open-weight or open-source releases associated with this launch.

---

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