DEV Community

AI OpenFree
AI OpenFree

Posted on

Darwin LLM Family Hits 1M+ Hugging Face Downloads: How VIDRAFT Built High-Performance Models on Minimal Resources

Darwin LLM Family Hits 1M+ Hugging Face Downloads: How VIDRAFT Built High-Performance Models on Minimal Resources

TL;DR: Korean AI startup VIDRAFT has released the Darwin family of open-source language models, surpassing 1 million cumulative downloads on Hugging Face just ~3 months after the first model dropped in April 2026. Built using an evolutionary model-merging methodology and a proprietary inference acceleration layer called VKAE, Darwin models punch well above their weight class on public benchmarks — and developers worldwide are already forking, quantizing, and deploying them.


What it is

Darwin is a family of open-source language models developed by VIDRAFT, a Korean early-stage AI startup. The model lineup spans multiple parameter scales, with publicly named variants including Darwin-398B-JGOS (a large-scale model) and Darwin-28B-Opus (a mid-size model). As of July 2026, the Darwin ecosystem on Hugging Face comprises:

  • 1,031,761 total cumulative downloads across all Darwin-related repositories
  • 269 repositories referencing Darwin models
  • 1,210 derivative/variant models created and redistributed by the global developer community
  • A single community-derived model, Darwin-9B-NEG, accounting for 560,000+ downloads on its own

The rapid community adoption — especially by developers who independently quantized or fine-tuned Darwin weights for redistribution — signals genuine grassroots traction rather than a coordinated marketing push. International AI communities and Chinese tech media have independently highlighted Darwin's technical approach.

VIDRAFT joins a short list of Korean organizations (alongside large-caps like LG, Naver, Upstage, and Kakao) that have crossed the 1M open-source download mark. As an early-stage startup, the milestone is notable.


How it works

VIDRAFT describes Darwin's training methodology as evolutionary: rather than pre-training a new model from scratch with massive compute, the team combines the strengths of existing models and iteratively advances them toward a next-generation checkpoint. Think of it as structured model merging with a selection pressure applied across generations — conceptually analogous to evolutionary algorithms applied to model capabilities.

On the inference side, VIDRAFT has developed VKAE, a proprietary compute acceleration technology. VKAE is applied at runtime to boost throughput on a single-machine deployment. According to the company, this achieves approximately 601 tokens per second on a single server, representing a 1.5× to 2× throughput improvement over baseline.

The company is extending Darwin and VKAE into on-device AI scenarios — edge deployments on smartphones, robotics platforms, and other resource-constrained hardware. A concrete example: VIDRAFT recently integrated its AI module into a Boston Dynamics Spot quadruped robot, enabling voice command and conversational interfaces running onboard.

The entire Darwin model family — including the 398B-parameter flagship — was reportedly developed using just 24 GPUs, underlining the efficiency philosophy behind the evolutionary approach.


Benchmarks & results

All figures below are sourced directly from VIDRAFT's public disclosures as reported by DBR:

Benchmark / Evaluation Model Result
GPQA Diamond (graduate-level science reasoning) Darwin-398B-JGOS 90.9% accuracy
GPQA Diamond Darwin-28B-Opus 88.89% accuracy
K-AI Blind Evaluation (Korean government-run) Darwin family 11 of top 20 models were Darwin variants
Polaris Drug Property Prediction Competition Darwin family 1st place in 14 out of 14 sub-tasks
Proprietary AI trap/hallucination avoidance test Darwin family 99.5% avoidance rate

The GPQA Diamond scores are particularly striking context: this benchmark is designed to challenge frontier models on PhD-level STEM reasoning. Darwin-28B-Opus's 88.89% at 28B parameters — and production on 24 GPUs — makes the efficiency story concrete.


How to try it

Darwin models are publicly available on Hugging Face. You can browse the official VIDRAFT organization page and search for Darwin model variants. The community has also published 1,210+ derived models under various names.

To download a Darwin model using the Hugging Face CLI:

pip install huggingface_hub
huggingface-cli download VIDRAFT/Darwin-28B-Opus
Enter fullscreen mode Exit fullscreen mode

Note: Replace VIDRAFT/Darwin-28B-Opus with the exact repository slug from the Hugging Face hub. Always verify the official organization page for the canonical model names before pulling weights.

The source article does not announce a public OpenAI-compatible API endpoint or a GitHub repository with training code at this time. Check VIDRAFT's Hugging Face organization and official channels for the latest access options.


FAQ

Q: What makes the evolutionary methodology different from standard fine-tuning or continued pre-training?
A: Instead of extending a single base model with additional training data, VIDRAFT's approach systematically combines capability signals from multiple existing models and applies an iterative selection process — pushing the merged result toward higher benchmark performance across generations. This reduces the compute required versus training from scratch.

Q: Is VKAE open-source, and can I use it with my own models?
A: Based on current public information, VKAE is a proprietary VIDRAFT technology and has not been released as an open-source library. Its availability as a standalone developer tool has not been announced. Watch VIDRAFT's official channels for updates.

Q: Are Darwin models multilingual, or Korean-focused?
A: CEO Minsik Kim explicitly frames Darwin as enabling sovereign AI — models adapted to each country's environment. While the Korean government benchmark performance is highlighted, the global developer adoption and community derivatives suggest broad multilingual utility. Specific language coverage details should be confirmed against the model cards on Hugging Face.


Originally reported by DBR (2026-07-13) — source article.

Top comments (0)