CL-SN55 941 —
GPT-61S 885 —
GPT-56T 861 —
GPT-6S 825 —
CL-OP55X 818 -0.1%
MUSE-SPK 793 —
QWEN-38X 779 —
GPT-6A 778 —
CL-OP5H 711 -0.1%
CL-OP5X 709 -0.1%
KIMI-K3X 699 -0.1%
CL-OP46H 694 -0.1%
CL-FAB5H 692 -0.1%
CL-OP55H 692 -0.1%
CL-OP47H 683 -0.1%
GEM-37FH 648 -0.2%
GEM-38FH 633 -0.2%
INKL 531 —
CL-OP47 529 -0.2%
GPT-56S 529 -5.9%
GEM-3P 493 —
CL-OP46 488 —
CL-OP48 481 —
GEM-31P 453 -0.2%
GPT-6L 436 —
CL-SN55 941 —
GPT-61S 885 —
GPT-56T 861 —
GPT-6S 825 —
CL-OP55X 818 -0.1%
MUSE-SPK 793 —
QWEN-38X 779 —
GPT-6A 778 —
CL-OP5H 711 -0.1%
CL-OP5X 709 -0.1%
KIMI-K3X 699 -0.1%
CL-OP46H 694 -0.1%
CL-FAB5H 692 -0.1%
CL-OP55H 692 -0.1%
CL-OP47H 683 -0.1%
GEM-37FH 648 -0.2%
GEM-38FH 633 -0.2%
INKL 531 —
CL-OP47 529 -0.2%
GPT-56S 529 -5.9%
GEM-3P 493 —
CL-OP46 488 —
CL-OP48 481 —
GEM-31P 453 -0.2%
GPT-6L 436 —
← Back to feed

VIDRAFT Darwin-180B-RSI Claims Open-Weight GPQA Diamond Lead at 94.4%

Korean startup VIDRAFT released Darwin-180B-RSI, a 180B open-weight reasoning model published on Hugging Face (FINAL-Bench/Darwin-180B-RSI). On AIhoster’s independent open-model benchmark board, it ranks first in two widely tracked evaluation categories.

Scores

BenchmarkScoreMethodology
GPQA Diamond94.4%Majority voting, up to 16 samples
MMLU-Pro88.1%12,032 test items
  • Thinking budget: up to 131,072 tokens
  • Precision: bf16
  • Weights: free on Hugging Face

The Methodology Caveat

GPQA Diamond at 94.4% is a striking number, but the methodology matters. The standard way frontier models are evaluated on GPQA Diamond — and the basis for comparisons between Claude Opus 5, GPT-6 Astra, and Gemini 4 Argon — uses single-pass evaluation. Majority voting over 16 samples increases accuracy mechanically: each additional sample gives the model more chances to land on the correct answer.

No direct single-pass result for Darwin-180B-RSI has been published. The 94.4% figure cannot be compared directly to standard leaderboard scores without the same harness.

VIDRAFT has not published results on swebench.com’s Verified leaderboard, Chatbot Arena, or Vals Index. AIhoster is an aggregator that runs its own evaluation suite.

What Darwin-180B-RSI Is

Darwin-180B-RSI is a dense 180B parameter reasoning model — not a sparse MoE architecture. At 180B active parameters per forward pass, it sits above the compute range of most flash-tier MoE models (3–25B active) but is still deployable on multi-GPU setups. A bf16 checkpoint at 180B requires approximately 360 GB of GPU memory.

The RSI suffix likely refers to the training objective or reinforcement schedule; VIDRAFT has not published a technical report.

Context for Open-Weight Reasoning

Among open-weight models with published GPQA Diamond numbers:

  • DeepSeek V4-Pro: 96.4% SWE-bench, strong across reasoning evals
  • GLM-5.3: 95.4% SWE-bench Verified (vals.ai)
  • Kimi K3: 93.4% SWE-bench Verified

GPQA Diamond at 94.4% via majority voting is plausible for a 180B dense model with extended thinking, though independent replication on a standard harness is needed before the claim holds against frontier peers.