DEV Community

AI OpenFree
AI OpenFree

Posted on

VIDRAFT's Darwin-398B-JGOS Takes #1 on Hugging Face Open LLM Leaderboard for Scientific Reasoning

#darwin398bjgos #huggingfaceleaderboa #scientificreasoningl #openweightmodel

VIDRAFT's Darwin-398B-JGOS Takes #1 on Hugging Face Open LLM Leaderboard for Scientific Reasoning

TL;DR: VIDRAFT, a Korean Pre-AGI AI startup, has released Darwin-398B-JGOS, a 398-billion-parameter model purpose-built for scientific reasoning that has claimed the top position on the Hugging Face Open LLM Leaderboard. This is a significant milestone for open-weight scientific reasoning models, and developers interested in cutting-edge reasoning capabilities will want to track this model's availability on Hugging Face.

What it is

Darwin-398B-JGOS is VIDRAFT's large-scale scientific reasoning model, carrying a 398-billion parameter count that places it firmly in the frontier model tier. Key facts from the source:

  • Model name: Darwin-398B-JGOS
  • Developer: VIDRAFT (비드래프트), a Korean Pre-AGI AI startup
  • Primary domain: Scientific reasoning — the model is explicitly designed and evaluated for tasks that require structured, multi-step scientific inference rather than general-purpose conversation or code generation alone
  • Leaderboard result: Ranked #1 on the Hugging Face Open LLM Leaderboard as of the reporting date (2026-06-17)
  • Model scale: 398B parameters, situating it alongside other frontier-class open-weight models

The "JGOS" suffix in the model name likely signals a specific variant or training configuration, though VIDRAFT has not publicly elaborated on what the acronym denotes beyond what is in the press coverage.

How it works

At a conceptual level, Darwin-398B-JGOS follows the general paradigm of large-scale transformer-based language models fine-tuned and/or trained with a strong emphasis on scientific reasoning tasks. While VIDRAFT has not publicly disclosed architectural internals or training hyperparameters, a few high-level points can be inferred from the model's positioning:

  • Scientific reasoning focus: Rather than optimizing purely for general language benchmarks, the model appears to be trained with a deliberate emphasis on domains requiring logical chains of thought — think physics, chemistry, mathematics, and biology-style problem solving
  • Scale as a lever: At 398B parameters, the model operates in a regime where emergent reasoning behaviors are well-documented in the research literature; this scale is consistent with competitive performance on multi-step inference tasks
  • Open-weight orientation: The model's presence on the Hugging Face leaderboard implies it has been submitted through the standard open evaluation pipeline, suggesting the weights are or will be accessible to the community rather than being API-only

No internal training details, loss functions, learning rates, or infrastructure specifics have been disclosed, and this article will not speculate beyond the above.

Benchmarks & results

The headline result reported by 데이터넷 is straightforward:

  • Hugging Face Open LLM Leaderboard: #1 ranking in the scientific reasoning category as of 2026-06-17

The source article does not provide granular per-benchmark scores (e.g., specific scores on GPQA, SciEval, MATH, or other individual benchmarks), so no numerical breakdown is available at this time. Developers looking for task-level scores should check the model's Hugging Face model card and the leaderboard entry directly, where detailed per-task metrics are typically published as part of the official evaluation pipeline.

Qualitatively, reaching #1 on the Hugging Face leaderboard — a competitive, community-audited benchmark environment — is a meaningful signal that Darwin-398B-JGOS outperformed other publicly evaluated open-weight models on the scientific reasoning axes the leaderboard measures.

How to try it

The source article confirms the model is associated with Hugging Face (given the leaderboard context), but specific access instructions — such as a confirmed model repository slug, download command, or API endpoint — have not been publicly detailed in the available press coverage.

If and when VIDRAFT makes the model publicly available, typical access patterns for Hugging Face-hosted models would look like:

# Generic pattern — verify the actual repo name on Hugging Face before running
huggingface-cli download <vidraft-org>/<model-repo-name>
Enter fullscreen mode Exit fullscreen mode

Developers should:

  1. Watch the VIDRAFT organization page on Hugging Face for the official model repository
  2. Check VIDRAFT's GitHub for any inference scripts or quantized variants
  3. Follow VIDRAFT's official channels for announcements about OpenAI-compatible API access, if offered

Do not use any endpoint or model identifier not confirmed by VIDRAFT directly — none are available to publish safely at this time.

FAQ

Q: Is Darwin-398B-JGOS open-weight, or is it API-only?
A: Based on the leaderboard context (Hugging Face Open LLM Leaderboard rankings typically require open weight submissions), the model is likely available as open weights, but VIDRAFT has not yet published explicit download or licensing instructions in the coverage available. Monitor the VIDRAFT Hugging Face organization page for the official release.

Q: What kinds of scientific reasoning tasks is this model targeting?
A: The model is described as a scientific reasoning model, which generally covers structured multi-step inference across STEM domains — mathematics, physics, chemistry, and biology-style problem solving. Specific supported task types and prompt formats will be detailed in the official model card once published.

Originally reported by 데이터넷 (2026-06-17) — source article.

Top comments (0)