VIDRAFT's Darwin-398B-JGOS Hits 90.9% on GPQA Diamond — What ML 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 scored 90.9% on GPQA Diamond — a graduate-level science reasoning benchmark considered highly challenging even for expert humans. This puts Darwin-398B-JGOS among a very small group of models at the frontier of expert-level reasoning, and signals that VIDRAFT is a serious player in the competitive large-model space worth tracking.
What it is
Darwin-398B-JGOS is VIDRAFT's flagship large language model, a dense or mixture-style model at the 398-billion-parameter scale. The model is developed by VIDRAFT (비드래프트), a Korean AI startup operating under a Pre-AGI research mission — meaning the company is explicitly oriented toward building systems capable of progressively more general and expert-level reasoning.
Key facts from the source:
- Model name: Darwin-398B-JGOS
- Developer: VIDRAFT (비드래프트), South Korea
- Scale: 398 billion parameters
- Headline result: 90.9% on the GPQA Diamond benchmark
GPQA Diamond is a particularly demanding evaluation: it consists of multiple-choice questions written by domain experts (primarily in biology, chemistry, and physics) and is specifically designed so that even PhD-level specialists in the relevant fields answer correctly only around 65–70% of the time on average. Crossing the 90% threshold on this benchmark is, as of mid-2026, a meaningful frontier-level result.
How it works
VIDRAFT has not publicly disclosed the full architectural or training details of Darwin-398B-JGOS, so the following is a high-level conceptual framing based on what is publicly known about models at this scale and capability level:
- Scale-driven reasoning: At 398B parameters, the model has sufficient capacity to encode broad factual knowledge across scientific domains — a prerequisite for performing well on expert-level Q&A benchmarks like GPQA Diamond.
- Instruction and reasoning alignment: Models achieving high GPQA Diamond scores typically undergo significant post-training work — including reinforcement learning from human or AI feedback — to sharpen multi-step scientific reasoning and reduce hallucination on verifiable factual questions.
- JGOS suffix: The "JGOS" designation in the model name likely indicates a specific variant, fine-tune, or training configuration within the Darwin model family, though VIDRAFT has not publicly elaborated on what this suffix represents.
The core takeaway for engineers: this is a large-scale model where careful post-training alignment, not just raw parameter count, appears to be doing meaningful work to push reasoning performance to frontier levels.
Benchmarks & results
The single headline benchmark reported in the source:
| Benchmark | Score |
|---|---|
| GPQA Diamond | 90.9% |
Context for engineers:
- GPQA Diamond is widely regarded as one of the hardest publicly available reasoning benchmarks for LLMs as of 2025–2026.
- Human expert baseline on GPQA Diamond is roughly in the 65–70% range; scores above ~85–87% are considered frontier-class as of early 2026.
- A 90.9% result, if reproducible under standard evaluation conditions, places Darwin-398B-JGOS in competitive territory with the top publicly benchmarked models globally.
No additional benchmark scores (e.g., MATH, HumanEval, MMLU) were reported in the source article. Developers should treat the single reported number as a headline claim pending independent reproduction.
How to try it
The source article does not publicly announce a Hugging Face repository, GitHub release, or API endpoint for Darwin-398B-JGOS at the time of this writing. Access details — if and when they become available — would most likely appear through:
- VIDRAFT's official channels (website or announcements)
- Hugging Face Hub under a VIDRAFT organization page
- An OpenAI-compatible API endpoint, which would be consistent with how Korean AI labs have typically offered developer access
Until VIDRAFT makes an official public release announcement, access to Darwin-398B-JGOS is not confirmed as publicly available. Watch VIDRAFT's official communications for release details.
FAQ
Q: How does 90.9% on GPQA Diamond compare to other leading models?
A: As of mid-2026, scores in the 88–92% range on GPQA Diamond represent the frontier of publicly reported results. A 90.9% score is competitive with top-tier models from major labs, though direct apples-to-apples comparison requires identical evaluation harness and prompt formatting — details VIDRAFT has not yet fully published.
Q: Is Darwin-398B-JGOS open-weights or proprietary?
A: The source article does not specify. No open-weights release or license information has been publicly announced as of the reporting date. Developers should check VIDRAFT's official channels for updates on model availability.
Q: What is the "JGOS" in the model name?
A: VIDRAFT has not publicly defined the JGOS designation. It likely denotes a specific variant or configuration within the Darwin model family, but no official explanation has been released.
Originally reported by 네이트 (2026-06-16) — source article.
Top comments (0)