Korean AI startup VIDRAFT shows that smarter AI doesn't always require more training data.
TL;DR: VIDRAFT, a Seoul-based Korean AI startup, published a preprint in May 2026 (arXiv:2605.14386) introducing the "Darwin Family," a framework that improves AI model performance by recombining the weights of existing models rather than running additional training. The flagship model, Darwin-27B-Opus, scored 86.9% on the GPQA Diamond benchmark, ranking sixth among 1,252 evaluated models at the time — surpassing its fully trained parent model — without any gradient-based training whatsoever.
VIDRAFT, a Korean AI startup headquartered in Seoul, made waves in China's tech media this week after Tencent News spotlighted the company's "Darwin Family" framework — a model-merging approach that enables AI systems to gain new capabilities purely by recombining the internal weights of pre-existing models, with zero additional training required. The research was released as a preprint in May 2026 under the identifier arXiv:2605.14386.
The central idea behind Darwin Family draws an analogy to animal breeding: just as a livestock expert can produce offspring that inherit the best traits of two parents without engineering an entirely new animal from scratch, VIDRAFT's researchers propose merging two AI models — one strong generalist base model and one reasoning-specialized model — to produce a hybrid that outperforms either parent.
The framework's headline result is Darwin-27B-Opus, which achieved an 86.9% score on GPQA Diamond, a rigorous scientific reasoning benchmark. At the time of evaluation, this placed the model sixth out of 1,252 models assessed — and, critically, above its "father" model, which had been trained through conventional full-training pipelines. Darwin-27B-Opus reached this level without undergoing any gradient-based training of its own.
The Darwin framework rests on three interlocking mechanisms:
MRI Importance Scoring. Borrowing the metaphor of a medical scan, this component analyzes every parameter layer in a model across two dimensions: static statistical properties (such as entropy, variance, and magnitude) and probe-response behavior, which measures how differently a layer activates when processing reasoning tasks versus casual conversation. Each dimension contributes equally to a final importance score, which then informs how much weight each parent model should contribute at that specific layer during merging.
Evolutionary Search and the Genome. Because MRI scoring is an approximation, Darwin adds an evolutionary search phase. Each candidate merge configuration is encoded as a 14-dimensional "genome" vector controlling variables such as global blending ratios, attention-module ratios, feed-forward-network ratios, embedding-layer ratios, sparsification densities for each parent model, blending ratios for six sequential layer blocks, an MRI trust parameter, and a merge-algorithm switching weight. The algorithm generates candidate genomes, evaluates their performance, retains the strongest, and iteratively mutates and recombines them — mirroring natural selection — until it converges on a high-performing configuration.
MRI Trust Fusion. A single parameter, τ (tau), governs the balance between following the MRI diagnostic recommendations and deferring to the evolutionary search's autonomous exploration. Crucially, τ is itself discovered by the evolutionary algorithm rather than set manually. Across all tested model sizes, τ consistently converged between 0.35 and 0.55, indicating that both information sources contribute meaningfully and neither alone is sufficient.
A fourth notable capability is the Architecture Mapper, which allows Darwin to merge models with different internal architectures — something most existing merge methods cannot do. A representative experiment, Darwin-4B-Genesis, successfully combined a Transformer-based model with a Mamba-based model, producing a hybrid that outperformed both parents on specific reasoning tests, again without retraining.
Training frontier AI models is extraordinarily expensive, often costing millions to tens of millions of dollars per run. The dominant assumption in the field has been that better performance demands bigger training budgets. VIDRAFT's Darwin Family challenges that assumption directly.
The research also builds on a well-documented tension in AI development: instruction fine-tuning and reinforcement learning can improve a model's surface-level accuracy on known task types while simultaneously degrading its deeper, generalized reasoning ability. Darwin sidesteps this trade-off by working entirely within the pre-trained weight space, where reasoning circuitry is already embedded, rather than layering on post-training that can distort it.
The cross-architecture merging capability is particularly forward-looking. As the AI field diversifies beyond Transformer-only designs, the ability to mix architectural lineages could prove as valuable as the training-free performance gains themselves.
Q: What is VIDRAFT's Darwin Family framework?
A: Darwin Family is a model-merging system developed by Korean AI startup VIDRAFT that combines the weights of two existing AI models — without any retraining — to produce a new model that can outperform either parent on reasoning benchmarks.
Q: How did Darwin-27B-Opus perform on benchmarks?
A: Darwin-27B-Opus achieved 86.9% on the GPQA Diamond scientific reasoning benchmark, placing sixth among 1,252 evaluated models at the time of publication, and surpassing its fully trained parent model — all without gradient-based training.
Q: Can Darwin merge AI models with different architectures?
A: Yes. Darwin's Architecture Mapper component is designed to align and merge models that have different internal structures. VIDRAFT demonstrated this with Darwin-4B-Genesis, a hybrid of a Transformer-based model and a Mamba-based model that outperformed both parents on targeted reasoning tasks without retraining.
Source: 텐센트 뉴스 (중국) (2026-05-21) — original article