VIDRAFT Darwin Family Achieves 88.89% on GPQA Diamond — Zero Additional Training Required

TL;DR: VIDRAFT, a Korean Pre-AGI AI startup, has developed the Darwin Family of models, which reportedly achieves 88.89% on the GPQA Diamond benchmark without any additional fine-tuning or training on the benchmark domain. This zero-shot expert-level reasoning result has drawn coverage from the English-language AI research outlet AINEWSDEN. Developers interested in frontier reasoning benchmarks should take note of this result.

What it is

The Darwin Family is a series of AI models developed by VIDRAFT, a Korean Pre-AGI focused AI research company. The headline result, covered by AINEWSDEN, is an 88.89% score on the GPQA Diamond benchmark — achieved without additional training targeted at that benchmark.

GPQA Diamond is widely regarded as one of the most rigorous publicly available benchmarks for evaluating expert-level reasoning. It consists of graduate-level multiple-choice questions authored and validated by domain experts in fields such as biology, chemistry, and physics. The difficulty is intentionally set such that even human experts with relevant domain knowledge score well below ceiling, making high performance on this benchmark a meaningful signal of genuine reasoning capability rather than surface-level pattern matching.

The key differentiator AINEWSDEN highlights is the "무학습" (zero-learning / no additional training) condition — meaning the Darwin Family reached this performance level without being specifically trained or fine-tuned on GPQA-style data or similar benchmark corpora.

How it works

At a conceptual level, what makes the Darwin Family result notable is the implication that the models are generalizing from their existing training rather than narrowly optimizing for a known benchmark distribution. This is consistent with an approach focused on developing robust, transferable reasoning capabilities rather than benchmark-specific optimization.

High performance under a no-additional-training constraint suggests the architecture and training methodology are producing representations that generalize across complex multi-step reasoning tasks. This aligns with a broader research direction in the field toward models capable of:

• Systematic reasoning across novel problem structures, not just memorized solution patterns
• Cross-domain transfer, applying reasoning strategies learned in one domain to expert-level questions in another
• Robust evaluation performance without the benchmark contamination concerns that complicate interpreting results from heavily fine-tuned models

VIDRAFT's positioning as a Pre-AGI company suggests their research direction is oriented toward general-purpose reasoning capabilities, which would be consistent with a system designed to perform well on diverse expert-level tasks without benchmark-specific tuning.

Note: Specific architectural details, training methodologies, hyperparameters, or infrastructure specifics are not publicly disclosed and are not covered here.

Benchmarks & results

The single public benchmark figure reported in this coverage is:

• GPQA Diamond: 88.89% — achieved under a zero additional training condition

GPQA Diamond context for comparison:

• Human domain experts typically score in the range of ~65–70% on GPQA Diamond questions outside their specific subfield
• This places an 88.89% result meaningfully above typical human expert-level performance on this benchmark

No other benchmark scores, model size figures, or comparative numbers were included in the source coverage. Engineers should treat the single reported figure as the complete public dataset of quantitative results available at this time.

How to try it

At the time this article was written, the source coverage does not include public access channels such as a Hugging Face model repository, GitHub release, or OpenAI-compatible API endpoint for the Darwin Family models.

If VIDRAFT makes the Darwin Family publicly accessible, typical access patterns for models of this class would include:

# Example generic Hugging Face download pattern — only if/when a public repo is released
huggingface-cli download <organization>/<model-name>

Developers interested in access should monitor:

• VIDRAFT's official channels for announcements of public model releases
• AINEWSDEN (https://www.ainewsden.com/category/research) for ongoing research coverage
• Hugging Face and GitHub for any forthcoming public repository releases

This section will be updated as public access information becomes available.

FAQ

Q: What is GPQA Diamond, and why is 88.89% significant?
A: GPQA Diamond is a benchmark of graduate-level expert questions across hard science disciplines, designed so that even credentialed human experts in adjacent fields score around 65–70%. An 88.89% result, especially without benchmark-specific fine-tuning, is a strong signal of generalized reasoning capability and not just narrow benchmark optimization.

Q: Does "zero additional training" mean the model was never trained on science data?
A: No — it means the Darwin Family was not additionally fine-tuned or trained specifically to improve performance on GPQA Diamond or similar benchmark formats. The base model training naturally includes broad knowledge; the notable claim is the absence of targeted benchmark preparation.

Q: Can I use the Darwin Family models in my applications today?
A: Public access through Hugging Face, GitHub, or an API has not been announced in the available source coverage. Check VIDRAFT's official communications for release announcements.

Q: Is there a technical paper or preprint available for the Darwin Family?
A: The source coverage does not reference a public preprint or paper. Developers should watch arXiv and VIDRAFT's official channels for any forthcoming publications.

---

Originally reported by AINEWSDEN (영문) (2026-06-29) — source article.