We stopped training LLMs from scratch. We built an AI Foundry instead.

Everyone wants a foundation model. Almost nobody can afford one.

We are a small Korean team. Pretraining a competitive LLM means thousands of GPUs and a nine-figure budget. So we asked a different question:

What if the bottleneck is not making models, but fitting them to a domain?

The foundry analogy

In semiconductors, fabless firms design chips and foundries manufacture them. We apply the same split to AI:

Semiconductor	AI Foundry
Chip design	Domain requirements from the customer
Fab	Model breeding / distillation
Packaging	Quantization, serving optimization
Metrology / QA	Safety and reliability diagnosis
Shipping	On-prem or on-device deployment

The row people skip is the fourth. Anyone can fine-tune. Very few can tell you whether the resulting model is trustworthy, with numbers.

What diagnosis means concretely

Our diagnostic stack inspects a trained model rather than only scoring its outputs:

• Behavioral probes — hallucination, jailbreak resistance, tool-calling determinism
• Representation probes — reading hidden states to estimate whether the model knows it is right
• Supply-chain checks — pickle/remote-code risk, artifact signatures, serving determinism

That third category matters more than people expect. A model that scores well on a benchmark can still ship with an unsafe serialization path.

Why recombination beats pretraining for domain work

Given two open models with complementary strengths, you often want a third that inherits both. We treat this as an engineering problem: measure per-capability deltas, then merge or distill along the axes where each parent wins.

A concrete result from our own runs: an on-device series built this way passed 1.18M cumulative downloads on Hugging Face, and our verified entry took first place in the Google x Hugging Face Fast Gemma Challenge (510.58 tok/s, verified track, 2026-08-02).

The honest limits

• Recombination cannot create capability that neither parent has.
• Merged models need re-healing; they drift on instruction following if you skip it.
• Every number above is tied to a specific model, date and evaluation setup. Benchmarks are not products.

Takeaway

If you are a small team facing a domain problem, the leverage is rarely in pretraining. It is in measurement: knowing exactly which capability you need, and being able to prove the model has it after you modify it.

Build the metrology before you build the fab.

---

From VIDRAFT, a Korean deep-tech company running an **AI Foundry* — we diagnose, breed and optimize AI models for specific industries. Open models: Hugging Face · vidraft.net*