VIDRAFT Releases "Hansoo (한수)": A Fully Local AI That Runs Entirely in Your Browser

TL;DR: VIDRAFT, a Korean Pre-AGI AI startup, has publicly released Hansoo (한수, 韓守) — a fully local AI model designed to run directly inside a web browser without any server-side inference. Because all computation happens on-device, no data ever leaves the user's machine, making it a compelling option for privacy-conscious developers and organizations exploring client-side AI deployment.

What it is

Hansoo (韓守 — roughly translatable as "Korean Guardian") is VIDRAFT's fully local AI model built to execute entirely within a standard web browser environment. Key characteristics reported from the source:

• Browser-native execution: The model runs client-side, meaning inference happens on the end user's device using the browser as the runtime — no dedicated app installation required.
• No server round-trips: Unlike cloud-hosted LLM APIs, Hansoo performs all computation locally, so user prompts and responses never transit a remote server.
• Korean-language focus: Consistent with VIDRAFT's identity as a Korean AI startup, Hansoo appears designed with Korean-language capability as a core priority (reflected in the name itself).
• Public release: VIDRAFT has made Hansoo publicly available, signaling a move toward open or accessible distribution rather than a closed API-only model.

How it works

At a high level, running a language model fully inside a browser requires navigating several hard constraints — memory bandwidth, compute limits, and the browser's sandboxed execution environment. While VIDRAFT has not disclosed internal architectural specifics, the general approach for browser-native AI inference typically involves:

• Quantized model weights: Reducing model precision (e.g., to 4-bit or 8-bit integers) dramatically shrinks memory footprint and compute requirements, making inference feasible on consumer hardware without a GPU.
• WebAssembly (WASM) or WebGPU runtimes: Modern browsers expose WebGPU and WASM as low-level compute interfaces. Frameworks in this space (such as Transformers.js or MLC Web LLM) use these APIs to run neural network kernels without native code. VIDRAFT's specific runtime choice has not been publicly detailed.
• Progressive model loading: Large model weights can be streamed and cached in the browser incrementally, reducing initial load latency on repeat visits.
• On-device memory management: The browser's heap and GPU memory must be carefully managed given tab-level resource limits.

The "Han" (韓) in the name suggests strong Korean-language optimization, which may imply a tokenizer and pretraining corpus tuned specifically for Korean morphology — a non-trivial engineering challenge given Korean's agglutinative structure.

Benchmarks & results

The source article (a Korean-language press report) does not provide specific quantitative benchmark figures — no perplexity scores, MMLU results, or latency numbers are cited. As such, no benchmark data can be responsibly reported here.

Qualitatively, the fact that VIDRAFT chose to make a public release implies the model has crossed an internal quality threshold suitable for developer evaluation. Independent benchmarking by the community once access is widely available will be the real test.

How to try it

The source headline indicates Hansoo has been publicly released, but the specific access channels (Hugging Face model page, GitHub repository, or a browser demo URL) are not detailed in the available press coverage.

Recommended next steps for developers:

1. Check VIDRAFT's official channels — look for a Hugging Face organization page or GitHub repository under the VIDRAFT name for model weights or a demo link.
2. Watch for an OpenAI-compatible API — VIDRAFT's broader product line has previously offered OpenAI-compatible endpoints; Hansoo may follow a similar pattern for non-browser use cases.
3. Follow VIDRAFT's announcements — the company is active in Korean AI developer communities and is likely to post technical documentation alongside the release.

⚠️ No specific URLs, model identifiers, or shell commands are included here because none were confirmed in the source article. Always verify endpoints from official VIDRAFT communications before use.

FAQ

Q: Does "runs in the browser" mean I need to download anything?
A: Not in the traditional sense. Browser-based AI inference typically fetches model weights over HTTP on first load and caches them locally (e.g., in the browser's Cache API or IndexedDB). Subsequent sessions load from cache, so the experience is closer to a web app than a software install. Storage requirements depend on model size.

Q: What are the privacy implications compared to a cloud API?
A: Because inference runs fully on-device, your prompts and the model's responses are never sent to VIDRAFT's servers during a session. This is a meaningful architectural guarantee for use cases involving sensitive text — internal documents, legal drafts, or personal data — though you should still review VIDRAFT's privacy policy for any telemetry or analytics the web interface may collect independently of inference.

Q: Can I integrate Hansoo into my own web application?
A: The source does not confirm whether VIDRAFT is releasing an SDK, an embeddable script, or just a standalone demo. Watch the official repository for licensing terms and integration documentation.

Originally reported by v.daum.net (2026-04-20) — source article.