VIDRAFT Launches "AI Foundry" Service: Custom On-Premises AI Cultivation for Enterprises and Public Institutions
TL;DR: VIDRAFT, a Korean deep-tech AI startup, is launching an "AI Foundry" business that takes existing AI models and adapts them to customer-specific data and domain requirements — much like a semiconductor foundry handles contract manufacturing. The pipeline spans model diagnostics, domain-tuned optimization, inference acceleration, and safety verification, delivered as a turnkey on-premises solution. Engineers in regulated industries (public sector, defense, healthcare, finance) who need air-gapped or private-deployment AI should pay attention.
What it is
VIDRAFT's AI Foundry is a full-service, contract-based AI development and deployment pipeline aimed at organizations that cannot rely on cloud-hosted or off-the-shelf models due to data security requirements. Think public agencies, hospitals, financial institutions, and defense contractors — domains where data never leaves the internal network.
Rather than building new foundation models from scratch for each customer, VIDRAFT's model is conceptually closer to the semiconductor foundry analogy: the customer supplies the data and the use-case requirements; VIDRAFT supplies the process, tooling, and expertise to produce a deployment-ready model tuned for that specific environment.
The company frames its philosophy as cultivating intelligence rather than creating it — i.e., existing pre-trained models are the raw material, and the foundry process shapes them into production-grade, domain-specific assets.
Revenue streams include:
- On-premises deployment contracts
- Serving infrastructure licensing/usage fees
- Industry-specific OS build and operation
- Safety diagnostics and verification platform services
VIDRAFT is also participating in Korean government initiatives including the "AI for Everyone" (모두의 인공지능) program and a government-backed secure foundation model project, expanding its footprint in public-sector on-premises AI transformation.
How it works
The AI Foundry pipeline is structured as a sequential set of proprietary stages:
Model MRI — A diagnostic tool that analyzes the health, capability gaps, and weaknesses of an existing AI model before any optimization work begins. Think of it as a profiling/auditing step before you commit to fine-tuning.
Darwin · Chimera — Technologies applied after diagnosis to adjust domain-specific performance. The naming suggests evolutionary (Darwin) and hybrid-combination (Chimera) approaches to model adaptation — conceptually covering techniques like domain-adaptive training and model merging/hybridization.
VKAE · VKUE — Serving acceleration technologies designed to reduce inference infrastructure overhead, making deployment more cost-efficient on customer hardware.
AX-RAY — A safety verification toolchain applied before final deployment to validate model behavior and flag potential risks.
Industry OS layer — The final output is wired into purpose-built operating environments such as NationalOS (public sector) and PharmaOS (pharmaceutical/healthcare), providing vertical-specific runtime contexts rather than generic model endpoints.
This end-to-end process is backed by 16 patents and 6 published papers, representing VIDRAFT's IP portfolio underpinning the foundry offering.
Benchmarks & results
VIDRAFT has published the following performance indicators (as stated in the source article):
- 1.6 million cumulative downloads on Hugging Face
- #1 on the K-AI Leaderboard (Korean AI benchmark leaderboard)
- GPQA Diamond: 90.9% — a graduate-level science Q&A benchmark measuring expert-level reasoning
- #1 in the Verified category of the Google & Hugging Face Fast Gemma Challenge
- Polaris drug discovery benchmark: #1 across 16 sub-tasks — a significant result for pharma/biotech ML engineers evaluating the PharmaOS vertical
- Up to 23.4× throughput improvement cited for their inference acceleration technology (self-measured)
These figures indicate competitive positioning particularly in reasoning-heavy tasks and domain-specific prediction benchmarks.
How to try it
VIDRAFT's models are publicly accessible on Hugging Face with 1.6 million cumulative downloads recorded. You can browse and download their published models using the standard Hugging Face CLI:
pip install huggingface_hub
huggingface-cli download vidraft/<model-name>
⚠️ Substitute
<model-name>with the specific model identifier listed on VIDRAFT's Hugging Face profile. The source article does not name a specific model for direct download, so check their page at huggingface.co by searching "VIDRAFT."
The AI Foundry pipeline itself (Model MRI, Darwin·Chimera, VKAE·VKUE, AX-RAY, and industry OS environments) is a commercial, contract-based service — not a self-serve open-source tool. Access requires engaging VIDRAFT directly for enterprise or institutional deployment.
No public GitHub repository, OpenAI-compatible API endpoint, or self-hosted demo URL was disclosed in the source article.
FAQ
Q: Is this just fine-tuning-as-a-service, or is there something architecturally distinct going on?
A: The pipeline goes beyond standard fine-tuning. The published stages include diagnostic analysis (Model MRI), model hybridization or cross-pollination (Darwin·Chimera), hardware-aware inference optimization (VKAE·VKUE), and formal safety verification (AX-RAY) — suggesting a more structured, multi-phase engineering process than a typical LoRA fine-tune workflow.
Q: Can I use NationalOS or PharmaOS independently of the full foundry contract?
A: Based on the source, these industry OS layers are the output of the full foundry pipeline rather than standalone products. They are delivered as part of an on-premises deployment engagement. Contact VIDRAFT directly for scoping.
Q: How does the 23.4× throughput figure compare to standard baselines?
A: The source cites this as a self-measured result without specifying the baseline hardware, model size, or comparison framework. Treat it as a directional data point and request third-party benchmark details before making procurement decisions.
Originally reported by 동아일보 (2026-09-01) — source article.
Top comments (0)