VIDRAFT's Darwin LLM Family Crosses 1M Hugging Face Downloads in 3 Months — Transformer + Mamba Architecture, 90.9% GPQA Diamond

#darwinllm #huggingface #openweightmodel #mambaarchitecture

VIDRAFT's Darwin LLM Family Crosses 1M Hugging Face Downloads in 3 Months — Transformer + Mamba Architecture, 90.9% GPQA Diamond

TL;DR: VIDRAFT, a Korean Pre-AGI AI startup, has announced that its Darwin family of evolutionary language models surpassed 1 million cumulative downloads on Hugging Face just three months after its April 2026 debut. The model line supports both Transformer and Mamba architectures, achieves 90.9% on the GPQA Diamond scientific reasoning benchmark, and is designed around an OpenAI-compatible API. If you're building on open-weight LLMs — especially in quantized or on-device contexts — Darwin is worth a close look.

What it is

Darwin is VIDRAFT's flagship open-weight language model family, framed by the company as an evolutionary approach to model development: combining the strengths of heterogeneous model architectures to push performance under resource-constrained conditions.

Key facts from the source:

First public release: April 2026
Cumulative Hugging Face downloads: 1.03 million+ as of July 2026
Total Darwin-family repositories on Hugging Face: 269
Quantized variant models created by the community: 1,210 distinct variants
Single breakout derivative: Darwin-9B-NEG alone accounts for 560,000+ downloads
Architecture support: Both Transformer and Mamba — an unusual dual-architecture stance that several international tech outlets flagged as a notable technical differentiator
Flagship models mentioned: Darwin-398B-JGOS and Darwin-28B-Opus

The majority of the download milestone was driven not by official VIDRAFT releases, but by community-led redistribution and quantization — a strong signal of organic developer adoption.

How it works

At a conceptual level, VIDRAFT describes Darwin's development philosophy as evolutionary model merging: selectively combining capabilities from distinct model lineages rather than training a single monolithic model from scratch. This approach is positioned as a way to achieve high benchmark performance without proportionally scaling compute.

The architecturally notable aspect is simultaneous support for Transformer and Mamba backbones. While most open-weight models commit to one paradigm, Darwin's design appears to treat architecture selection as a variable — enabling deployment targets ranging from large-scale inference to on-device and physical AI applications.

VIDRAFT also ships a proprietary inference acceleration layer called VKAE, which the company says enables 601 tokens per second on a single GPU. The system is exposed through an OpenAI-compatible API, meaning any application already targeting the OpenAI REST interface can point at Darwin without client-side changes.

The company is actively extending Darwin into vision-language model (VLM) territory and physical AI — including a demonstrated integration with Boston Dynamics' Spot quadruped robot, where a Darwin-based VLM module powers voice command and conversational interaction.

Benchmarks & results

All figures below are sourced directly from VIDRAFT's public announcement as reported by 중앙일보:

Model	Benchmark	Score
`Darwin-398B-JGOS`	GPQA Diamond (scientific reasoning)	90.9%
`Darwin-28B-Opus`	GPQA Diamond	88.89%

Additional reported outcomes:

Korean government K-AI Blind Evaluation: 11 of the top 20 models in the ranking were from the Darwin family
Polaris drug property prediction competition: Darwin-family models claimed 1st place across 14 categories
VKAE inference throughput: 601 tokens/sec on a single GPU (single-GPU environment, as stated by VIDRAFT)

GPQA Diamond is a graduate-level scientific reasoning benchmark widely used to stress-test frontier models. Scores above 85% place a model in highly competitive territory relative to publicly ranked systems.

How to try it

Darwin models are publicly available on Hugging Face. The large community of 1,210+ quantized variants means you can likely find a version sized for your hardware — from full-precision flagship models down to heavily quantized formats suitable for consumer GPUs or edge devices.

Browse and download models:

# Search for Darwin models on Hugging Face
huggingface-cli search vidraft/darwin

# Download a specific model (substitute the actual repo name you find)
huggingface-cli download <repo-id>

OpenAI-compatible API access is noted as part of VIDRAFT's commercialization design. If you already use the OpenAI Python SDK or curl-based requests, the integration pattern would follow the standard base URL override:

from openai import OpenAI

client = OpenAI(
    base_url="<VIDRAFT_API_ENDPOINT>",  # check VIDRAFT's official docs for the current endpoint
    api_key="<YOUR_API_KEY>",
)

⚠️ Public API endpoint details and pricing are not disclosed in this source article. Check VIDRAFT's official site or Hugging Face organization page for the latest access instructions.

FAQ

Q: What makes Darwin different from other open-weight Korean LLMs?
A: The dual Transformer + Mamba architecture support, the evolutionary model-merging development philosophy, and the VKAE inference acceleration layer set it apart technically. On the ecosystem side, crossing 1M downloads in three months — driven primarily by community quantization — is unusual for a non-US-origin model.

Q: Can I run a quantized Darwin model on a consumer GPU?
A: Almost certainly yes, depending on the variant. With 1,210+ community-quantized derivatives already on Hugging Face, formats like GGUF and GPTQ are likely well represented. Search the Darwin family repositories for quantized variants matching your VRAM budget.

Q: Is the OpenAI-compatible API available publicly right now?
A: The source article states the system is designed for OpenAI-compatible API commercialization, but does not publish a specific public endpoint. Verify current availability directly with VIDRAFT's official channels.

Q: What is VKAE?
A: VKAE is VIDRAFT's proprietary inference acceleration technology. The company reports it achieves 601 tokens/sec on a single GPU. Architectural or implementation details beyond this are not publicly disclosed in the source.

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