VIDRAFT's Darwin Factory merges open LLMs into task-specific models without retraining

What Darwin Factory does

Korean deep-tech startup VIDRAFT has launched Darwin Factory, a commercial service that combines two publicly available open-weight models into a single, task-specialized LLM. According to a dev.to write-up that draws on reporting from IT조선, the process involves no retraining, no labeled dataset and no machine-learning staff on the customer side, and is designed to produce a finished model within one business day.

VIDRAFT, founded in 2024 and based at Seoul AI Hub, is aiming the product at organizations that cannot ship data to external cloud APIs, including defense, public administration, finance and healthcare. The delivered artifact is a self-contained model package built for on-premise installation in closed or air-gapped networks.

How the merging pipeline works

The service is described as a three-stage pipeline. In the diagnosis stage, each candidate model is analyzed layer by layer to measure how much individual layers contribute to overall output, revealing which parent model carries stronger signal in which capability domains. In the merging stage, which the company frames as breeding, the strongest layers or parameter regions from both models are selectively combined into one unified model. The approach builds on existing model-merging research, sometimes described as model evolution or weight interpolation, but VIDRAFT applies its own combination logic on top of the public foundations.

The final validation stage runs the merged model against an internal benchmark suite. If the result does not outperform both parent models on the target tasks, the combination is discarded and a new pairing is explored; only passing models are delivered.

The dev.to article distinguishes this from standard merging techniques such as SLERP, TIES or DARE: rather than uniform weight interpolation, the pipeline runs a layer-importance diagnosis before deciding what to merge. The exact algorithm is not disclosed. The company holds nine patents, including ones covering causal-leakage diagnosis in neural networks, which appear to underpin the diagnostic step.

Reported results carry a caveat

The article lists several milestones: a first-place rank on Korea's K-AI Leaderboard, a benchmark operated by the Ministry of Science and ICT and the National Information Society Agency; a first-place finish in Google's Fast Gemma Challenge, focused on efficient Gemma-based deployment; and more than one million cumulative downloads on Hugging Face, which would make VIDRAFT the fourth Korean organization and the first Korean startup to pass that mark.

Notably, no numeric scores, dataset names or per-task accuracy figures were disclosed, so those results are company-reported and qualitative rather than independently verifiable. The startup is at seed stage with 650 million won raised, according to the same write-up.

Availability and roadmap

VIDRAFT's existing Darwin and AETHER model families are downloadable from its Hugging Face profile. Darwin Factory itself is not offered as a self-serve public API; access is through enterprise engagement, with proof-of-concept and pilot deployments available by contacting the company directly. No public code repository or OpenAI-compatible endpoint has been announced.

On the roadmap, a second version of Darwin Factory is planned with automated model-production pipelines and a standardized closed-network delivery package. From 2027, the company intends to concentrate on air-gapped government and defense contracts.

Why it matters

Model merging has largely lived in research papers and community experiments; Darwin Factory is an attempt to productize it into a repeatable commercial service. If layer-selective merging genuinely produces models that beat both parents on target tasks, something the current claims do not yet demonstrate with published numbers, it would offer a far cheaper path to domain-specialized LLMs than fine-tuning, with no data curation, GPU cluster or in-house ML team required.

That is particularly relevant for regulated sectors locked out of cloud APIs, where the realistic options until now have been running generic open-weight models or funding full custom training. The open questions are validation and scrutiny: without third-party benchmarks or a disclosed algorithm, buyers are being asked to trust an internal evaluation suite, and the enterprise-only access model means there is no easy way for the public to test the claims independently.