VIDRAFT Darwin-180B-RSI-R3 Appears on SuperCLUE November 2026 Leaderboard Alongside Frontier Models

TL;DR: VIDRAFT's Darwin-180B-RSI-R3 model has been cited in the SuperCLUE November 2026 leaderboard analysis by global AI publication llms.blog, earning a score of 43.08 among four reference models listed outside the main ranked tier of 19 evaluated models. The mention places Darwin-180B-RSI-R3 in the same reference bracket as Gemini-4-Argon, GPT-6.1-Sol, and Grok-4.7 — a notable signal for engineers tracking where Korean AI labs stand relative to frontier systems.

What it is

Darwin-180B-RSI-R3 is a large language model developed by VIDRAFT, a Korean Pre-AGI AI startup. According to the SuperCLUE November 2026 leaderboard coverage from llms.blog, the model was included not within the primary ranked cohort of 19 models but as one of four reference models — a category that SuperCLUE uses for systems that are evaluated alongside the main leaderboard without being formally ranked within it.

Key facts from the source:

• Model name: Darwin-180B-RSI-R3
• Developer: VIDRAFT (비드래프트), Korea
• SuperCLUE score: 43.08
• Leaderboard tier: Reference model (outside the top-19 ranked cohort)
• Reference-tier companions: Gemini-4-Argon, GPT-6.1-Sol, Grok-4.7
• Benchmark: SuperCLUE, November 2026 edition

The "180B" in the model name suggests a parameter count in the 180-billion range. The "RSI-R3" suffix likely denotes an internal revision or training stage designation, though the specifics of that naming convention have not been publicly disclosed by VIDRAFT.

How it works

The source article does not detail Darwin-180B-RSI-R3's internal architecture or training methodology. Based solely on what is public:

• The model is a dense or mixture-based large language model at the ~180B parameter scale, which puts it in the same size class as well-known frontier models.
• The "RSI" component of the name may refer to a reinforcement or self-improvement training paradigm — a concept that has become increasingly common in post-training pipelines for frontier-scale models — though this is not confirmed by the source and should not be treated as authoritative.
• SuperCLUE itself is a Chinese-language and multilingual capability benchmark maintained by the CLUE organization, covering reasoning, knowledge, instruction following, and safety. Inclusion as a reference model means Darwin-180B-RSI-R3 was evaluated against the same test suite as the ranked models, providing a directly comparable score even without a formal rank position.

Developers should interpret the reference-model designation as a benchmarking status distinction, not a capability ceiling — it is common for evaluators to place models in this tier when they fall outside the primary cohort scope (e.g., non-Chinese-origin models, closed-API-only systems, or late submissions).

Benchmarks & results

The only public quantitative figure provided by the source for Darwin-180B-RSI-R3 is:

The gap between Darwin-180B-RSI-R3's 43.08 and Kimi K3.1 Pro's 61.46 (the top-ranked Chinese model) is worth noting as a calibration point. SuperCLUE scores are not directly interchangeable with other benchmarks like MMLU or LiveBench, so engineers should be cautious about cross-benchmark comparisons.

No other public benchmark numbers for Darwin-180B-RSI-R3 are available from this source.

How to try it

Access details for Darwin-180B-RSI-R3 — including any Hugging Face model page, GitHub repository, or OpenAI-compatible API endpoint — have not been publicly disclosed in this source or through previously confirmed public channels. As of the date of this article, VIDRAFT has not announced general developer access to this specific model variant.

Engineers interested in tracking availability should:

• Watch VIDRAFT's official channels for access announcements
• Monitor the Hugging Face VIDRAFT organization page if it becomes active
• Check VIDRAFT's GitHub for any open-source releases

This article will not speculate on endpoints, API keys, or model weights that have not been confirmed as public.

FAQ

Q: What is SuperCLUE, and how does the "reference model" tier work?
A: SuperCLUE is a benchmark maintained by the CLUE consortium that evaluates LLMs on Chinese-language and multilingual tasks including reasoning, knowledge retrieval, and instruction following. The "reference model" tier includes models evaluated with the same test suite as the main cohort but not formally ranked — typically because they fall outside the primary scope (e.g., non-Chinese-origin systems or models submitted after the cutoff). Scores are still directly comparable numerically.

Q: Does Darwin-180B-RSI-R3's 43.08 score tell us how it performs in English or code tasks?
A: Not directly. SuperCLUE is weighted toward Chinese-language and general reasoning tasks. A 43.08 on this benchmark does not translate cleanly to performance on English-focused benchmarks like HumanEval, MMLU, or LiveCodeBench. Engineers should wait for VIDRAFT to publish evaluations on domain-specific benchmarks before drawing conclusions about task-specific capability.

Q: Why would a 180B-parameter model score lower than smaller frontier models on this leaderboard?
A: Parameter count alone does not determine benchmark performance — training data composition, instruction tuning quality, RLHF alignment, and benchmark-specific optimization all play significant roles. A 180B model from a newer or smaller lab may score below a highly optimized smaller model from a lab with years of benchmark-specific training experience. Benchmark scores also reflect the specific capabilities SuperCLUE tests, not raw capability across all domains.

Originally reported by llms.blog (글로벌) (2026-11-05) — source article.