VIDRAFT's Darwin LLM Family Hits 1M Hugging Face Downloads — Evolutionary Training on a Lean GPU Budget

TL;DR: VIDRAFT, a Korean Pre-AGI startup, has released its open-source Darwin language model family on Hugging Face, crossing 1 million cumulative downloads roughly three months after the first model dropped. Darwin is built with an evolutionary model-merging methodology rather than full pretraining from scratch, yielding strong benchmark scores at unusually low compute cost. If you're evaluating efficient open-weight LLMs for sovereign AI, on-device deployment, or scientific reasoning tasks, this family is worth a look.

What it is

Darwin is a family of open-weight large language models developed by VIDRAFT and distributed via Hugging Face. Key facts from the source:

• Cumulative downloads: 1,031,761 as of the reporting date (July 13, 2026), ~3 months after the first model release in April 2026.
• Ecosystem footprint: 269 repositories and 1,210 derivative/quantized variants created by the global developer community, not by VIDRAFT directly. A single community derivative, Darwin-9B-NEG, alone accounts for 560,000+ downloads.
• Model size range: The family spans at least two flagship sizes: a large variant (Darwin-398B-JGOS) and a mid-size variant (Darwin-28B-Opus).
• Designed for sovereign AI scenarios: VIDRAFT explicitly positions Darwin as a foundation for country- or domain-specific AI deployments — an increasingly competitive niche as organizations seek locally controllable models.
• On-device extension: VIDRAFT is extending the technology toward on-device AI use cases, including a recent integration with a Boston Dynamics Spot quadruped robot for voice-command and conversational interaction.

How it works

VIDRAFT takes a conceptually different approach compared to training a model from scratch:

• Evolutionary methodology: Rather than pretraining on raw data from zero, VIDRAFT combines the strengths of existing well-performing models and iteratively evolves them toward the next generation. Think of it as principled model merging or model-soup techniques applied as a systematic development loop — you start with strong priors from the existing model landscape and refine from there.
• Compute efficiency focus: The headline benchmark results were produced using only 24 GPUs — a deliberately constrained budget for a lab aiming at 90%+ GPQA scores. The methodology is explicitly designed to extract maximum capability per GPU-hour.
• VKAE inference acceleration: VIDRAFT ships a proprietary compute-acceleration layer called VKAE (details are internal) that the company reports achieves 601 tokens/second on a single machine — approximately 1.5× to 2× faster than the baseline without VKAE. This is particularly relevant for on-device and edge deployment targets.

No unpublished hyperparameters, training schedules, or architectural internals are available from the source.

Benchmarks & results

All numbers below come directly from the Dong-A Ilbo source article:

Benchmark / Evaluation	Model	Result
GPQA Diamond (scientific reasoning)	Darwin-398B-JGOS	90.9%
GPQA Diamond (scientific reasoning)	Darwin-28B-Opus	88.89%
Korean Government K-AI Blind Evaluation	Darwin family	11 of top 20 models are Darwin variants
Polaris Drug Property Prediction Competition	Darwin family	1st place in 14 categories
Internal AI trap-question evasion rate	Darwin family	99.5%
Inference throughput (single machine, VKAE)	—	601 tokens/second

The GPQA Diamond score for the 28B-parameter class model is notably competitive — for context, GPQA Diamond is considered a strong proxy for graduate-level scientific reasoning and is one of the harder public benchmarks in circulation.

How to try it

Darwin models are publicly available on Hugging Face. Because the community has generated 1,210+ derivative repositories, you can search for the base models and community variants directly on the platform:

# Search and download base models via the Hugging Face CLI
pip install huggingface_hub
huggingface-cli download VIDRAFT/Darwin-28B-Opus

⚠️ Note: The exact Hugging Face organization slug and all available model cards are on the VIDRAFT Hugging Face page. Browse the community derivatives (including Darwin-9B-NEG and quantized GGUF/GPTQ variants) by searching Darwin VIDRAFT on huggingface.co. No public GitHub repo URL or OpenAI-compatible API endpoint was disclosed in the source article at time of writing.

---

FAQ

Q: Is Darwin a fine-tune of an existing open model, or is it something architecturally novel?
A: VIDRAFT describes an "evolutionary methodology" that combines strengths of existing models rather than training from scratch — this is closer to advanced model merging / evolutionary model composition than to a novel architecture. The source does not claim a new transformer variant; the novelty is in the training and composition strategy and the VKAE inference layer.

Q: How does a 28B model score 88.89% on GPQA Diamond with only 24 GPUs?
A: The efficiency gain comes from the evolutionary approach — by bootstrapping from already-capable models, VIDRAFT avoids the enormous compute cost of pretraining from random initialization. The 24-GPU figure refers to the development compute budget, not inference hardware requirements.

Q: Can I use Darwin for commercial projects?
A: License terms are not specified in the source article. Check the model card on Hugging Face directly for the applicable license before production use.

Q: Is the VKAE acceleration layer open source?
A: It is not described as open source in the source article. VKAE appears to be a proprietary inference optimization developed internally by VIDRAFT.

---

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