A 36-billion-parameter model created in under an hour is matching models ten times its size on graduate-level reasoning benchmarks.
TL;DR: VIDRAFT's research division VIDRAFT_LAB, operating under the FINAL-Bench project, has released Darwin-36B-Opus, a 36-billion-parameter AI model built using an evolutionary weight-recombination engine called Darwin V7 rather than conventional training. The model scored 88.4% on the GPQA Diamond benchmark, matching a 397-billion-parameter competitor and ranking third on the leaderboard as of August 23, 2026. The entire creation process reportedly took less than one hour on a single GPU.
VIDRAFT, the Korean Pre-AGI AI startup, made waves in the global AI community when Russian technology outlet vc.ru published an in-depth analysis on August 23, 2026, spotlighting Darwin-36B-Opus — a model that challenges a foundational assumption of modern AI development: that cutting-edge performance requires weeks of compute-intensive training from scratch.
Darwin-36B-Opus is the latest output of Darwin V7, an evolutionary engine developed by FINAL-Bench / VIDRAFT_LAB. Rather than undergoing conventional pre-training or fine-tuning on massive datasets, the model was created by taking two existing "parent" models and automatically recombining their weights to produce a higher-performing offspring.
The two parent models were Qwen/Qwen3.6-35B-A3B, which contributed its hybrid attention mechanism and a 256-expert routed topology, and hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled, which brought chain-of-thought reasoning behavior distilled from Claude Opus 4.6. According to the developers, the resulting model inherits the mother's distilled reasoning patterns while preserving the structural precision of the father's expert architecture.
The headline benchmark result is an 88.4% score on GPQA Diamond, a rigorous 198-question evaluation covering graduate-level physics, chemistry, and biology. That figure places Darwin-36B-Opus:
On the GPQA Diamond leaderboard dated April 23, 2026, Darwin-36B-Opus ranked third overall, behind only TNSA/NGen-4-Pro at 91.1% and TNSA/NGen-4 at 90.1%.
Technically, the model uses a Mixture-of-Experts (MoE) architecture. Its 36 billion total parameters include only 3 billion active parameters at inference time — just 8 of its 256 experts are engaged per forward pass — making it far leaner to run than its benchmark scores might suggest. Additional specifications include a 262,000-token context window, BF16 precision, multilingual support across nine languages (English, Chinese, Korean, Japanese, German, French, Spanish, Russian, and Arabic), and an Apache 2.0 open-source license permitting free commercial use.
The conventional playbook for building a frontier AI model involves enormous GPU clusters, terabytes of training data, weeks of compute time, and budgets that can run into the tens or hundreds of millions of dollars. Darwin-36B-Opus represents a direct challenge to that paradigm.
By treating existing open-weight models as a "genetic pool" and systematically recombining their weights through Darwin V7, VIDRAFT_LAB produced a model that competes with architectures more than ten times its size — in under an hour on a single GPU. If this approach proves reproducible and generalizable, it could dramatically lower the cost barrier to developing high-capability AI systems.
The open Apache 2.0 license compounds the significance. Enterprises and researchers do not need special permissions or commercial agreements to deploy Darwin-36B-Opus, which contrasts with the more restrictive licensing of many frontier-class models. Combined with only 3 billion active parameters at runtime, the model is accessible to organizations that cannot afford to operate hundred-billion-parameter systems.
The multilingual coverage — including Korean, Russian, Arabic, and Japanese alongside the major Western European languages — also signals an intentional push toward international deployability rather than English-first design.
Q: What is Darwin-36B-Opus and who made it?
A: Darwin-36B-Opus is a 36-billion-parameter AI language model developed by FINAL-Bench / VIDRAFT_LAB, the research arm of Korean AI startup VIDRAFT. It was built using an evolutionary engine called Darwin V7 that recombines weights from two parent models rather than training from scratch.
Q: How does the Darwin V7 evolutionary engine work?
A: Darwin V7 takes two existing open-weight models, automatically recombines their parameters, and produces a new model that inherits beneficial characteristics from both parents. According to the developers, the full process can complete in less than one hour on a single GPU.
Q: How does Darwin-36B-Opus perform compared to much larger models?
A: On the GPQA Diamond benchmark, Darwin-36B-Opus scored 88.4%, equaling a 397-billion-parameter model and outperforming several other large models including a 744-billion-parameter system, while ranking third on the overall GPQA Diamond leaderboard as cited by the source.
Source: vc.ru (러시아) (2026-08-23) — original article