Darwin-180B-RSI 登上 AIhoster 的 MMLU-Pro 和 GPQA Diamond 排行榜首位

Darwin-180B-RSI Tops MMLU-Pro and GPQA Diamond on AIhoster's Independent Open-Model Benchmark Board

TL;DR: VIDRAFT's Darwin-180B-RSI, a 180-billion-parameter open model, has claimed the #1 position on both the MMLU-Pro and GPQA Diamond leaderboards as tracked by Russian benchmark aggregator AIhoster — scoring 88.1% and 94.4% respectively, outperforming models with parameter counts an order of magnitude larger. Developers interested in frontier reasoning capabilities in an open-weight format should take note.

What it is

Darwin-180B-RSI is a 180B-parameter open-weight language model developed by VIDRAFT (비드래프트), a Korean Pre-AGI AI startup. It is published under the Hugging Face repository identifier FINAL-Bench/Darwin-180B-RSI.

AIhoster is an independent Russian benchmark aggregation site that collects publicly reported evaluation results from model cards, papers, and leaderboards — without recomputing or normalizing scores into a composite index. Each benchmark entry is tied to an exact repository ID and a specific evaluation configuration, making comparisons methodologically transparent.

As of the AIhoster update on 2026-10-19, Darwin-180B-RSI holds:

• #1 on MMLU-Pro — evaluated on all 12,032 test items
• #1 on GPQA Diamond — evaluated on all 198 items

Notably, VIDRAFT has multiple Darwin-family models appearing in the GPQA Diamond rankings simultaneously: Darwin-397B-ZTC (#3 at 93.4%) and Darwin-398B-JGOS (#9 at 90.9%), suggesting an active, iterative development pipeline across different model scales and inference configurations.

How it works

At a conceptual level, Darwin-180B-RSI appears to be a large-scale transformer-based language model optimized for extended reasoning. A few things can be inferred from the publicly disclosed evaluation configurations:

• Extended thinking budget: The MMLU-Pro and GPQA evaluations both used a thinking budget of up to 131,072 tokens — consistent with a chain-of-thought or "reasoning model" architecture that allocates substantial token budget for intermediate reasoning steps before producing a final answer.
• Sampling strategy varies by benchmark: For MMLU-Pro, a single sample was drawn (temperature=1.0, top_p=0.95, top_k=20). For GPQA Diamond, majority voting over up to 16 samples was used under identical sampling parameters — a test-time compute scaling strategy that trades inference cost for accuracy on hard problems.
• Precision: The model runs in bf16, which is standard for large-scale inference on modern accelerators.
• Efficiency relative to parameter count: Achieving #1 on both benchmarks with 180B parameters — while outperforming models like DeepSeek-V4-Pro (1.6T parameters) and Kimi-K2.5 (1T parameters) on MMLU-Pro — suggests strong parameter efficiency, likely from architectural choices or training data curation rather than raw scale alone.

No architectural specifics, training hyperparameters, or dataset details are publicly disclosed in the source material.

Benchmarks & results

All figures below are sourced directly from AIhoster's leaderboard (updated 2026-10-19), which in turn cites the Darwin-180B-RSI model card.

MMLU-Pro (multi-domain expert knowledge; higher accuracy = better):

GPQA Diamond (expert-level physics, chemistry, biology; higher accuracy = better):

The GPQA result is particularly notable: Darwin-180B-RSI achieves 94.4% using majority voting over up to 16 samples, while the #3 model, Darwin-397B-ZTC (also a VIDRAFT model), achieves 93.4% with greedy decoding and no voting — a different inference regime entirely.

How to try it

The model is listed under the Hugging Face organization FINAL-Bench with the repository FINAL-Bench/Darwin-180B-RSI. If the weights are publicly available there, you can browse or download via the standard Hugging Face CLI:

huggingface-cli download FINAL-Bench/Darwin-180B-RSI

Access details beyond what is on the model card — including any API endpoints, licensing terms, or inference service availability — are not confirmed in the source material. Check the model card directly for the most current access instructions.

FAQ

Q: How does Darwin-180B-RSI beat models with 10× more parameters on MMLU-Pro?
A: The source doesn't explain internal architecture, but the evaluation configuration gives a clue: the model uses a 131,072-token thinking budget, suggesting extended chain-of-thought reasoning. Parameter efficiency at inference is a known benefit of well-optimized mixture-of-experts or distillation-based architectures, though VIDRAFT has not publicly confirmed which approach is used here.

Q: Are the GPQA and MMLU-Pro scores independently verified?
A: AIhoster explicitly states it does not recompute scores — it aggregates from public model cards and leaderboard sources, matched by exact repository ID. The source for Darwin-180B-RSI's scores is listed as its own model card. Independent third-party reproduction has not been confirmed in this reporting.

Q: Can I use Darwin-180B-RSI commercially?
A: Licensing terms are not specified in the source article. Check the model card at FINAL-Bench/Darwin-180B-RSI on Hugging Face for the applicable license before any production or commercial use.

Originally reported by AIhoster (러시아) (2026-10-18) — source article.