cd /news/artificial-intelligence/vidraft-s-darwin-180b-claims-12-1-ra… · home › topics › artificial-intelligence › article
[ARTICLE · art-171407] src=dev.to ↗ pub= topic=artificial-intelligence verified=true sentiment=↑ positive

VIDRAFT's Darwin-180B Claims 12 #1 Rankings Across 57 Hugging Face Official Leaderboards

Korean AI startup VIDRAFT published an audit of Hugging Face's officially tagged benchmarks claiming its Darwin-180B model, built on a Recursive Self-Improvement (RSI) mechanism, took 12 first-place finishes across 57 leaderboards spanning science, math, document parsing, law, agents and retrieval, with top-3 results on all 15 leaderboards it entered. The company ranked ahead of Moonshot AI (7), Z.ai (6), DeepSeek (3) and Alibaba Qwen (3) by number-one count in the November 17, 2026 API snapshot, and CEO Kim Min-sik attributed the cross-domain performance to RSI's domain-agnostic self-improvement loop. The model card is public on Hugging Face as FINAL-Bench/Darwin-180B-RSI, though no API endpoints or licensing details were disclosed.

recommendation

When citing leaderboard rankings, restrict the query to datasets tagged benchmark:official, count ties as shared first places, and publish the snapshot date so others can replicate it.

more in playbook: Evals and testing for LLM apps
by read4 min views5 publishedNov 17, 2026

TL;DR: Korean AI startup VIDRAFT has published a comprehensive audit of all Hugging Face officially certified benchmarks, showing their Darwin (다윈) series models topped 12 leaderboards across six distinct domains — science, math, document parsing, law, agents, and retrieval. The models achieved a top-3 finish on every single leaderboard they entered (15 of 15), making VIDRAFT the institution with the most #1 rankings globally in this snapshot. Engineers evaluating frontier reasoning models for research or production use cases now have a new data point worth examining.

VIDRAFT (비드래프트) is a Korean deep-tech AI company focused on large-scale language models and reasoning systems. Their flagship product line centers on Darwin (다윈), a self-improving model platform. The specific model highlighted in this analysis is Darwin-180B, which VIDRAFT describes as powered by their RSI (Recursive Self-Improvement) technology.

VIDRAFT's product ecosystem also includes:

The analysis in question audited all Hugging Face datasets carrying the benchmark:official tag — covering 63 benchmark datasets, 57 leaderboards, and 1,458 records — using data from the Hugging Face public API as of November 17, 2026. Ties were counted as shared first places for all tied parties.

VIDRAFT's CEO Kim Min-sik (김민식) attributes the cross-domain performance to the RSI (Recursive Self-Improvement) mechanism built into Darwin. At a conceptual level, RSI refers to a training or fine-tuning paradigm in which the model iteratively improves its own outputs — using its own generated responses as signals for subsequent learning rounds, rather than relying purely on external human-labeled data.

The claimed advantage of RSI is domain generality: because the self-improvement loop does not require domain-specific annotation pipelines, the same core mechanism can transfer performance gains across heterogeneous fields like graduate-level science, legal exams, mathematical competition problems, and structured document extraction. Whether the model reasons, retrieves, or classifies, the underlying improvement signal is model-generated and domain-agnostic.

This positions Darwin's RSI approach conceptually alongside other self-play and self-critique methods seen in recent frontier research, but VIDRAFT claims their implementation generalizes more reliably across benchmark categories than comparable systems from other institutions.

All numbers below are sourced directly from the November 17, 2026 Hugging Face official API snapshot:

#1 Finishes (12 total across 6 domains):

Near-top finishes:

Participation rate:

Country-level context:

Institution ranking by #1 count: VIDRAFT (12) → Moonshot AI (7) → Z.ai (6) → DeepSeek (3) → Alibaba Qwen (3)

Darwin-180B also reportedly claimed first place for open-source models in the Math & Science Reasoning category on SuperCLUE's November 2026 evaluation — a Chinese AI benchmarking organization.

VIDRAFT has made Darwin-180B's model card publicly available on Hugging Face. You can view it at:

FINAL-Bench/Darwin-180B-RSI on Hugging Face.

huggingface-cli repo info FINAL-Bench/Darwin-180B-RSI

No additional public API endpoints, GitHub repositories, or OpenAI-compatible API details were disclosed in the source material. Check the Hugging Face model page directly for the latest access instructions, weights availability, and licensing terms.

Q: How was the leaderboard count methodology determined — could it inflate VIDRAFT's numbers?

A: VIDRAFT used only Hugging Face datasets tagged benchmark:official, covering 63 datasets and 57 leaderboards as of November 17, 2026. Ties were counted as first place for all tied parties (e.g., AIME 2026 is a shared #1). The methodology is explicitly described as a snapshot of public API data, not a proprietary evaluation — anyone can replicate the query against the same tag.

Q: Does a 100% top-3 rate on 15 leaderboards mean Darwin-180B is a general-purpose frontier model?

A: Not necessarily. VIDRAFT submitted to 15 of 57 official leaderboards, meaning there are 42 leaderboards where Darwin-180B has no recorded entry. The 100% top-3 rate reflects strong performance in the domains VIDRAFT chose to benchmark, not universal coverage. For the most contested boards (MMLU-Pro at 163 models, HLE at 104 models, SWE-bench Verified at 78 models), Darwin-180B placed first on GPQA Diamond but second on MMLU-Pro and has no listed record on HLE or SWE-bench Verified.

Q: Is RSI the same as RLHF or DPO?

A: Based on the source, RSI (Recursive Self-Improvement) is described as a self-improvement loop where the model iterates on its own outputs. The source does not equate it to RLHF, DPO, or any specific published technique. Treat it as a proprietary methodology until VIDRAFT publishes technical details.

Originally reported by 넷이즈 网易 (중국) (2026-11-17) — source article.

── more in #artificial-intelligence 4 stories · sorted by recency
── more on @vidraft 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/vidraft-s-darwin-180…] indexed:0 read:4min 2026-11-17 · —