VIDRAFT's Darwin-398B-JGOS Tops Global AI Scientific Reasoning Leaderboard

TL;DR: VIDRAFT, a Korean Pre-AGI AI startup, has announced that its Darwin-398B-JGOS model has reached the #1 position on a global leaderboard for AI scientific reasoning. This 398-billion-parameter model represents a significant milestone in large-scale reasoning systems aimed at scientific domains. Developers and researchers interested in frontier reasoning models should take note of this benchmark result.

What it is

Darwin-398B-JGOS is VIDRAFT's large language model specifically designed and evaluated for scientific reasoning tasks. Key facts from the source:

• Model name: Darwin-398B-JGOS
• Parameter scale: 398 billion parameters — placing it firmly in the frontier LLM tier alongside other hundred-billion-plus scale models
• Developer: VIDRAFT (비드래프트), a Korean Pre-AGI AI startup
• Achievement: Ranked #1 on a global AI scientific reasoning leaderboard as of June 2026
• Domain focus: Scientific reasoning — a category that typically covers tasks such as mathematics, physics, chemistry, biology, and multi-step logical inference grounded in scientific knowledge

The "JGOS" suffix in the model name is specific to VIDRAFT's naming convention; the "Darwin" series name signals an intentional alignment with scientific discovery and systematic reasoning as core design goals.

How it works

At a conceptual level, models targeting scientific reasoning at this scale generally combine several well-understood approaches, and what is publicly known about VIDRAFT's direction is consistent with this:

• Scale-driven reasoning: At 398B parameters, the model operates in a regime where emergent multi-step reasoning capabilities become more reliable — critical for scientific problem-solving that requires chaining intermediate conclusions
• Reasoning-optimized training: Scientific reasoning benchmarks reward structured, verifiable outputs. Models optimized for this domain are typically trained with objectives that reinforce step-by-step derivation rather than surface-level pattern matching
• JGOS architecture framing: The "JGOS" designation suggests a specific model configuration or training methodology within VIDRAFT's Darwin series, though the precise architectural details are not disclosed in the source
• Leaderboard-driven evaluation: Reaching #1 on a global scientific reasoning leaderboard implies the model was evaluated against standardized, publicly recognized benchmarks covering multi-domain scientific tasks — the kind that require genuine knowledge synthesis, not retrieval shortcuts

It's worth emphasizing: VIDRAFT has not publicly disclosed internal training configurations, data mixtures, or architectural specifics beyond what is reported here.

Benchmarks & results

The source reports a #1 global ranking on an AI scientific reasoning leaderboard for Darwin-398B-JGOS. Specific numeric scores per sub-benchmark are not detailed in the available coverage. What can be stated qualitatively:

• The result positions Darwin-398B-JGOS above all other models currently tracked on the referenced leaderboard at the time of reporting (June 2026)
• Scientific reasoning leaderboards of this type typically aggregate performance across multiple domains — mathematics, formal logic, natural sciences — making a #1 overall ranking a broad signal of capability rather than narrow task specialization
• VIDRAFT frames this as a meaningful step toward their Pre-AGI research goals, suggesting the scientific reasoning domain is a core evaluation axis for the company's roadmap

No specific numeric scores (e.g., accuracy percentages, pass@k values) were available in the source article — this article will not fabricate them.

How to try it

The source article does not provide specific public access instructions — no Hugging Face repository link, GitHub URL, or API endpoint is mentioned in the available coverage.

If VIDRAFT follows patterns common among frontier model labs, access channels to watch would include:

• Hugging Face Hub: Search for VIDRAFT or Darwin-398B-JGOS on huggingface.co
• VIDRAFT's official channels: Monitor their announcements for API access, waitlists, or research partnership programs
• OpenAI-compatible API: If VIDRAFT offers an inference endpoint, it may follow the standard /v1/chat/completions interface pattern

⚠️ Do not attempt to construct endpoints or model paths speculatively. Wait for official release announcements before integrating into any pipeline.

FAQ

Q: What exactly is the "global scientific reasoning leaderboard" Darwin-398B-JGOS topped?
A: The source references a global AI scientific reasoning leaderboard without naming the specific benchmark suite or hosting organization. Watch VIDRAFT's official communications for the precise leaderboard citation and methodology.

Q: How does 398B parameters compare to other frontier models in practice?
A: 398B parameters puts Darwin-398B-JGOS in the same rough scale tier as other leading frontier models (e.g., mixture-of-experts and dense models in the 100B–400B+ range). At this scale, models generally exhibit stronger multi-step reasoning, better knowledge coverage across scientific domains, and improved instruction-following — all relevant to scientific reasoning tasks.

Q: Is this model open-weight or proprietary?
A: The source does not specify. No open-weight release or license has been announced in the available coverage. Follow VIDRAFT's official channels for updates on model availability.

Q: What is VIDRAFT's broader research direction?
A: VIDRAFT describes itself as a Pre-AGI AI startup, indicating that frontier reasoning capability — particularly in verifiable, structured domains like science — is central to their long-term research agenda. Darwin-398B-JGOS appears to be a significant waypoint on that path.

---

Originally reported by 전자신문 (2026-06-15) — source article.