VIDRAFT's Darwin-398B-JGOS Scores 90.9% on GPQA Diamond — What Engineers Need to Know
TL;DR: VIDRAFT, a Korean Pre-AGI AI startup, has released benchmark results for its large language model Darwin-398B-JGOS, which achieved 90.9% on GPQA Diamond — a graduate-level science reasoning benchmark that has become a key frontier indicator for cutting-edge LLMs. If you're tracking state-of-the-art reasoning models or evaluating alternatives for hard scientific Q&A tasks, this result is worth your attention.
What it is
Darwin-398B-JGOS is a large language model developed by VIDRAFT (비드래프트), a South Korean AI startup positioning itself in the Pre-AGI space. Key facts from the source:
- Model name: Darwin-398B-JGOS
- Scale: 398 billion parameters — placing it firmly in the frontier model tier alongside the largest publicly discussed LLMs
- Benchmark highlight: 90.9% accuracy on GPQA Diamond, a challenging multiple-choice benchmark composed of graduate-level questions in biology, chemistry, and physics, designed specifically to be difficult for non-expert humans and to stress-test genuine scientific reasoning in LLMs
- Developer: VIDRAFT (비드래프트), a Korean AI company focused on Pre-AGI research and development
GPQA Diamond is widely regarded as one of the harder evaluation suites in current LLM benchmarking circles because its questions are crafted by domain experts and validated to be resistant to surface-level pattern matching or web-memorization shortcuts.
How it works
The source article does not disclose architectural internals, training procedures, or implementation details of Darwin-398B-JGOS, and we won't speculate. At a conceptual level, what the benchmark result tells us:
- Strong scientific reasoning: A 90.9% GPQA Diamond score implies the model handles multi-step inference across STEM domains at a level that surpasses most non-expert humans on these tasks (human expert baseline on GPQA Diamond is typically cited around 65–70% in the research community)
- 398B parameter scale: Models at this parameter count typically leverage sophisticated approaches to knowledge capacity and generalization — the exact architecture and training methodology are not publicly confirmed at this time
- "JGOS" designation: The suffix in the model name likely indicates a specific variant, tuning configuration, or release version within VIDRAFT's Darwin model family, though the precise meaning has not been officially detailed in the source coverage
Developers should treat this as a capability signal: the model is optimized for tasks demanding rigorous, multi-domain scientific reasoning rather than purely conversational or instruction-following use cases.
Benchmarks & results
The single confirmed public benchmark result from the source:
| Benchmark | Score |
|---|---|
| GPQA Diamond | 90.9% |
- GPQA Diamond tests graduate-level reasoning in physics, chemistry, and biology via multiple-choice questions written and validated by domain experts
- 90.9% is a notably high score on this suite — for reference, GPQA Diamond has historically been used to differentiate frontier models precisely because it resists the performance inflation seen on older, more saturated benchmarks
- No additional benchmark figures (MMLU, HumanEval, MATH, etc.) are confirmed in the source article; the article centers specifically on this GPQA Diamond milestone
How to try it
The source article does not confirm public access to Darwin-398B-JGOS through Hugging Face, GitHub, an OpenAI-compatible API endpoint, or any other developer-facing channel at the time of reporting.
If you want to stay updated on access:
- Watch VIDRAFT's official channels for announcements about model releases or API access programs
- Check Hugging Face (
huggingface.co/vidraftor similar) for any public model drops — no confirmed repository URL is available as of this writing - Follow VIDRAFT's developer communications for API waitlists or partnership programs
Note to developers: Do not rely on any third-party endpoints claiming to serve Darwin-398B-JGOS until VIDRAFT officially confirms public access. We will not publish unverified connection strings or endpoints.
FAQ
Q: How does 90.9% on GPQA Diamond compare to other frontier models?
A: GPQA Diamond has become a key differentiator benchmark precisely because scores above ~85–90% are rare and signal genuine expert-level scientific reasoning. A 90.9% result places Darwin-398B-JGOS in competitive territory with the top tier of currently reported frontier LLMs. Exact cross-model comparisons depend on evaluation conditions (prompting strategy, sampling parameters, etc.), so always check whether methodologies are comparable before drawing hard conclusions.
Q: Is 398B parameters a dense model or a mixture-of-experts (MoE) architecture?
A: The source article does not specify the architectural design. The 398B figure may refer to total parameters (as in MoE) or active parameters in a dense setup — this distinction matters significantly for inference cost. Await official technical documentation from VIDRAFT for a definitive answer.
Q: What does "Pre-AGI" mean in VIDRAFT's positioning?
A: "Pre-AGI" is VIDRAFT's stated research orientation — it signals that the company is explicitly working toward artificial general intelligence as a long-term goal, with current models representing milestones on that path. It is a research direction label, not a claim about current capability.
Originally reported by 이데일리 (2026-06-16) — source article.
Top comments (0)