One column, honestly labelled

Start with what checks out. SuperCLUE is an external benchmark rather than a VIDRAFT-authored evaluation. Its November leaderboard uses a private test set, which limits direct training to disclosed test items but also prevents independent reproduction from the published materials. On that board, a 180B open-weight model from a small Korean lab outscored Gemini-4-Argon, DeepSeek-V4.1-Flash, Grok-4.7 and Qwen3.8-Max in mathematical and scientific reasoning. Its parent model, from which Darwin inherited all 512 routed experts untouched, scored 57.89 in the same column. Darwin's 73.68 is 15.79 percentage points above the 57.89 reported for its parent-model entry on the same SuperCLUE column. That is a substantial observed score difference on an external benchmark. It does not isolate the contribution of R3, establish statistical significance or show how much of the difference comes from earlier training rounds and evaluation settings.

The ranking language needs one qualifier. This is second place in one of five columns. On the overall index that combines all five domains, Darwin scores 43.08, and SuperCLUE's table gives it no rank at all: it sits in a reference group alongside Gemini-4-Argon, GPT-6.1-Sol and Grok-4.7, while the ranked list is topped by Kimi-K3.1-Pro at 61.46. VIDRAFT's model card describes Darwin as ninth among the 19 models when all displayed overall-index scores are sorted numerically. That is a derived ordering, not an official SuperCLUE overall rank. SuperCLUE places Darwin in its reference group and assigns it no rank in the overall table. The distinction matters: a sortable score is not necessarily an entry in the benchmark's ranked competition.

A one-item-sized gap, on the published pattern

The published math-and-science scores form an unusually regular ladder: 76.32, 73.68, 71.05, 71.05, 68.42, 68.42, 65.79 and 63.16, with the parent model at 57.89. Adjacent levels are separated by approximately 2.63 percentage points, with small differences caused by rounding. That spacing is consistent with a 38-item, equally weighted test, where each item would account for about 2.63 points. But SuperCLUE does not disclose the item count or scoring formula in the material available to the desk. The one-item interpretation is therefore an inference from the score pattern, not a confirmed description of the evaluation.

The ranking is official within this column; its apparent precision is harder to assess. Darwin trails GPT-6.1-Sol by 2.64 points and leads the two models tied immediately behind it by 2.63 points. If the column uses approximately 38 equally weighted items, those differences would each correspond to one item. Without the scoring methodology, the desk cannot establish that interpretation or estimate statistical significance. The published numbers support a narrow numerical lead, not a demonstrated capability gap.

What the self-improvement actually bought

Darwin's method is the part of this story most likely to matter in a year, and the R3 card documents it with rare precision. The model solves practice problems, keeps only the solutions that verify as correct, and trains on those. No human-written solutions, no teacher model. For R3, the raw material was 3,000 SuperGPQA questions the previous round had never trained on. The first round solved each one eight times, and only the boundary problems survived, the ones it answered correctly two to six times out of eight: 462 problems, yielding 714 correct solutions. Training touched the attention paths and shared experts, through a LoRA adapter that was then merged. The experts holding the parent's knowledge were not modified.

The measured return on that round is in the same card. On 1,000 held-out SuperGPQA questions that played no part in training or selection, R3 improved on R1 by 1.03 points on average, with a 95% confidence interval from +0.05 to +2.00. That interval barely excludes zero. The same card reports GPQA Diamond for all three generations under one matched protocol. Under the matched eight-sample, 32K-budget GPQA Diamond protocol, the parent scored 90.91 on majority vote, R3 scored 90.40 and R1 scored 89.90. The parent therefore had the highest reported majority-vote score in this comparison. The differences are small, and the card does not establish that they are statistically significant. R3's single-sample accuracy was 85.86, compared with 84.85 for the parent and R1, a difference of 1.01 percentage points. On a 198-question benchmark, that is approximately two question-equivalents, although the reported accuracy may aggregate multiple samples rather than represent one fixed set of two additional correct answers.

Set against that, the number on the model's banner, GPQA Diamond 94.44, comes from the first round under a different protocol: a majority vote over up to 16 samples with a thinking budget of 131,072 tokens. The perfect scores travelling with this family, 100 on AIME 2026 and 100 on HMMT February 2026, are majority-of-16 results as well; the card lists the mean single-sample accuracies beside them, 98.75 and 96.59. To the card's credit, none of this is hidden. It states that its numbers are self-measured, that majority-vote scores are system scores, and that the Hugging Face leaderboard values it compares against are self-reported by each publisher under differing settings. A reader who stops at the badges gets a different model from a reader who finishes the card.

Open weights, datacenter serving

The practical terms are the usual ones for a model of this size, plus one unusual one. In full precision the weights take about 336 GB, and the card's own serving example uses eight B200 GPUs. VIDRAFT also lists a 111 GB 4-bit GGUF build and reports that it can run with an 8 GB GPU and 32 GB of system memory, including a claimed CPU generation speed of 18–21 tokens per second and unchanged MMLU-Pro accuracy of 87.65%. Those are publisher-reported results, not independently verified measurements. A 111 GB weight file cannot fit entirely into the stated 40 GB of combined GPU and system memory. Running it under those conditions would therefore require some form of storage-backed access, offloading or another serving arrangement not established by the figures alone. The card's laptop claim needs a reproducible configuration before it can be treated as practical guidance. The license is the Qwen Community License 1.0, inherited from the parent.

One smaller disclosure is worth keeping. The model card also describes a confidence probe, ZTC, intended to estimate the likelihood of a correct answer before generation. VIDRAFT reports an AUROC of 0.64 on its own held-out split and characterizes the signal as coarse. That result shows limited discrimination in the reported evaluation; it does not establish calibrated probabilities or a reliable threshold for approving consequential actions. The probe is an experimental signal, not a validated safety gate. That is the right way to ship an early feature. It is also a useful calibration for the rest of the card: this is a lab that labels its weak numbers as weak. The strong ones still need their protocols read.

Sources

  1. [1] VIDRAFT / FINAL-Bench, Darwin-180B-RSI-R3 model card, Hugging Face (read in full 8 November 2026)Read source
  2. [2] VIDRAFT / FINAL-Bench, Darwin-180B-RSI model card, Hugging Face (read in full 8 November 2026)Read source
  3. [3] SuperCLUE, November 2026 general leaderboard (published 5 November 2026), as reproduced on the model cardsRead source
  4. [4] LLMs Blog, "SuperCLUE Nov 2026: Kimi K3.1 Pro Tops China Models" (independent reading of the SuperCLUE table, checked 5 November 2026)Read source
  5. [5] dev.to, "Darwin-180B-RSI-R3: VIDRAFT Releases Weights for Its Recursively Self-Improved 180B MoE Reasoning Model" (8 November 2026)Read source