Modelpublic

FINAL-Bench/Darwin-180B-RSI

sourceHugging Faceotherupdated 5d agoView on Hugging Face

121likes937downloads

Model Card

Darwin-180B-RSI

180B Mixture-of-Experts · vision-language · #1 on ten Hugging Face official leaderboards: nine by this model (MDPBench 83.65 · IFStruct 98.95 · AIME 2026 100 · HMMT Feb 2026 100 · GPQA Diamond 94.44 · MMLU-Pro 88.12 · MMMU-Pro 79.48 · LEXam 68.94 · LEXam-hard 45.72) and ExtractBench 90.29 by its next round R3 · self-improving

💻 Run it on your own machine — [POCKET-Darwin-180B-GGUF](https://huggingface.co/FINAL-Bench/POCKET-Darwin-180B-GGUF): the 4-bit GGUF of R3 (111 GB) runs on a laptop with an 8 GB GPU and 32 GB RAM, CPU only at 18–21 tok/s, a 128 GB mini PC or one DGX Spark — MMLU-Pro identical to BF16 (87.65%).

🥇 NEW: #1 on [MDPBench](https://huggingface.co/datasets/Delores-Lin/MDPBench) (83.65), multilingual document parsing across 17 languages, digital and photographed pages, ahead of Kimi-K3 (83.6) and the dedicated OCR models.

🥇 NEW: #1 on [IFStruct](https://huggingface.co/datasets/LiquidAI/ifstruct-v1.0) (98.95), the structured-output compliance benchmark (valid JSON/YAML that follows a requested schema), ahead of Agents-A1 (93.25) and gpt-oss-20b (91.95).

🥇 NEW: [Darwin-180B-RSI-R3](https://huggingface.co/FINAL-Bench/Darwin-180B-RSI-R3), the next self-improvement round of this model, is #1 on the [ExtractBench](https://huggingface.co/datasets/llamaindex/ExtractBench) leaderboard (90.29) and #3 on EvasionBench (77.83).

reasoning · MoE 512 experts · 262K long context · image + text · Korean + English · self-improvement · ZTC

The newest flagship of the Darwin family: #1 on MDPBench, IFStruct, AIME 2026, HMMT Feb 2026, GPQA Diamond, MMLU-Pro, MMMU-Pro, LEXam and LEXam-hard, and a model that gets better by learning from its own verified work.

🏆 Head-to-head with Chinese frontier models

Darwin-180B-RSI vs Chinese frontier models

Five leaderboards — full field

Scores as listed on the Hugging Face official benchmark leaderboards (self-reported by each model's publisher). 🥇 = #1 on that leaderboard. "—" = not reported.

Model	AIME 2026	GPQA Diamond	MMLU-Pro	MMMU-Pro	HMMT Feb 2026	LEXam	LEXam-hard
🧬 Darwin-180B-RSI (ours · 🇰🇷)	100 🥇	94.44 🥇	88.12 🥇	79.48 🥇	100 🥇	68.94 🥇	45.72 🥇
Inkling (Thinking Machines)	—	—	—	—	—	—	40.82
Kimi-K3 (Moonshot AI)	—	93.5	—	—	—	—	29.54
Kimi-K2.6 (Moonshot AI)	96.4	90.5	—	79.4	92.7	—	36.18
DeepSeek-V4-Pro (DeepSeek)	—	90.1	87.5	—	—	—	38.93
Qwen3.5-397B-A17B (Alibaba)	93.33	88.4	87.8	—	87.88	—	—
MiniMax-M2.1 (MiniMax)	—	80.81	88	—	—	—	—
GLM-5 (Zhipu AI)	95.83	86	86	—	86.36	—	—
Intern-S2-Preview (Shanghai AI Lab)	—	—	88	76.88	87.31	—	—
Step-3.5-Flash (StepFun)	96.67	83.5	84.4	—	86.36	—	—
DeepSeek-R1 (DeepSeek)	—	—	—	—	—	52.41	—
Qwen3-235B-A22B-Thinking-2507 (Alibaba)	—	—	—	—	—	48.19	—

This comparison covers open-weight models listed on the Hugging Face official benchmark leaderboards; closed API models are not included. Leaderboard values are each publisher's own reported numbers; settings (samples, voting, thinking budget) differ across models. Darwin-180B-RSI settings are listed in the evaluation protocol below.

🧬 The Darwin Family

Darwin is VIDRAFT's measurement-driven reasoning model family — 50+ official models, 400+ community derivatives, and now two places in the GPQA Diamond top 3 (Darwin-180B-RSI #1 · Darwin-397B-ZTC #3).

🧬 Darwin — evolve the parent, keep what works

Darwin treats a strong open model as a parent. It measures where the parent is weak, and strengthens exactly those parts — instead of re-training everything and risking what already works.

—Diagnose before you change. Every Darwin generation starts from a measured weakness map of the parent.
—Change little, precisely. Darwin modifies a small, targeted fraction of the network. Knowledge stored in the experts is preserved.
—Proven capability over new guesses. Earlier Darwin generations grafted the best-performing expert/FFN blocks from other strong models onto a base backbone; Darwin-180B-RSI adds a new ingredient — the model's own verified work.
—Measured, not claimed. Every change must beat the parent on held-out tests before it ships.

Model	Scale	GPQA Diamond
Darwin-9B-NEG	9B	84.3
Darwin-27B-Opus	27B dense	86.9
Darwin-36B-Opus	36B MoE	88.4
Darwin-28B-REASON	28B + DELPHI	89.39
Darwin-397B-ZTC	397B MoE (FP8)	93.43
Darwin-180B-RSI	180B MoE	94.44

Lineage

Role
Parent	`Qwen/Qwen3.8-Flash-Next`	180B MoE vision-language backbone · Qwen Community License 1.0
Darwin RSI	self-improvement on verified answers	the parent's own solutions, checked against verifiable answer keys, fed back as training signal
Preserved	512 routed experts · router · vision encoder	untouched — the parent's knowledge stays intact
ZTC	zero-token confidence readout	see below

📄 Darwin Platform & Research

—Darwin Family — MRI trust-weighted evolutionary merging for training-free scaling of language-model reasoning (arXiv:2605.14386)
—Placement Is Free, Composition Is Not — the Latin square as a provably-balanced construction for heterogeneous sequence-mixer stacks (2609.20269) — the AETHER architecture line
—FINAL Bench — VIDRAFT's measurement-driven evaluation framework (SSRN)
—Four-layer Pre-AGI roadmap — Darwin → AETHER → PROMETHEUS → HEPHAESTUS
—Collections: Darwin Family · ZTC Models — JEV ecosystems

🔁 RSI — a model that improves from its own work

Recursive self-improvement (RSI) is the core of this generation. Instead of distilling a bigger teacher, the model improves by learning from itself:

1.Solve — the model works through practice problems it has never seen in evaluation.
2.Verify — its answers are checked against verifiable references (answer keys, executable checks). Nothing unverified is learned.
3.Learn — it is re-trained on the reasoning that turned out to be correct.
4.Repeat — the improved model becomes the next solver.

What it bought in this release:

	Parent (Qwen3.8-Flash-Next)	Darwin-180B-RSI
Average reasoning length (MMLU-Pro)	4,320 tokens	3,833 tokens (−11 %)
MMLU-Pro accuracy	88.04 %	88.12 %

Same or better accuracy with shorter reasoning — cheaper and faster to serve. Practice sets are deduplicated against every evaluation set we report (8-gram overlap filter).

Next round: Darwin-180B-RSI-R3

Darwin-180B-RSI-R3 is the next RSI round, trained from this model with the same recipe. Its own leaderboard entries (submitted 2026-10-05):

Benchmark	R3	Leaderboard
ExtractBench (370 documents)	90.29, #1	llamaindex/ExtractBench
EvasionBench (16,726 questions)	77.83, #3	FutureMa/EvasionBench

Model-level RSI vs. harness-level RSI

Darwin-180B-RSI is Model-level RSI: the model itself (its weights) improves by learning only from its own solutions. No human-written solutions or reasoning traces are used; correctness is checked automatically. Harness-level RSI (e.g., Google's RRSI) improves the prompts, tools and workflow around a fixed model. It's like rewriting an employee's manual, while Model-level RSI is the employee getting smarter. The two are complementary.

🏛️ ZTC — it knows before it answers

Zero-Token Confidence (ZTC) reads the model's own internal state once, before generation, and returns the probability that the answer it is about to give is correct — no extra tokens, no second model.

json

{"answer": "...", "confidence": 0.93, "ztc_score": 1.84, "truncated": false}

Use it to gate actions: when confidence is low, do not call the tool, escalate, or answer "I don't know".

What ships with this model

File	Role
`handler.py`	one call returns the answer and its confidence as JSON
`ztc/ztc_probe_darwin180rsi.npz`	the ZTC readout for this model (final layer, last prompt token)
`ztc/usage.py`	minimal example

python

from handler import EndpointHandler
h = EndpointHandler("FINAL-Bench/Darwin-180B-RSI")   # local snapshot path
print(h({"inputs": "What is 17 * 23?"}))
# [{"answer": "...391...", "confidence": 0.97, "ztc_score": 2.1, "truncated": false}]

Same format as Darwin-397B-ZTC. The readout is fitted only on practice data that is disjoint from every benchmark reported here.

Status: early release. This probe reaches an AUROC of 0.64 on our held-out validation split: a coarse signal for routing and review, not a correctness guarantee. A retrained probe with more data will replace it. Details in `ztc/README.md`.

🏆 Results

Benchmark	Score	Setting	Leaderboard
GPQA Diamond (198)	94.44	majority vote over up to 16 samples · 131,072-token thinking budget	#1
MMLU-Pro (12,032)	88.12	single sample · 131,072-token thinking budget	#1
AIME 2026 (30)	100.0	majority vote over 16 samples (mean accuracy 98.75) · 131,072-token thinking budget	#1
HMMT Feb 2026 (33)	100.0	majority vote over 16 samples (mean accuracy 96.59) · 131,072-token thinking budget	#1
MMMU-Pro (vision, 1,730)	79.48	majority vote over 3 samples · 131,072-token thinking budget	#1
LEXam (law, MCQ 4-choice, 1,655)	68.94	majority vote over 4 samples (single sample 60.54 · mean 61.42) · 32,768-token thinking budget	#1
LEXam-hard (law, open-ended, 518)	45.72	single sample · 32,768-token thinking budget (60 truncated answers regenerated at 120K) · judged by DeepSeek-R1-0528 per the official eval.yaml	#1
IFStruct (structured output, 2,000 prompts)	98.95	single run · official harness defaults (temperature 0, 16,000 max tokens) · thinking on	#1
MDPBench (multilingual document parsing, 2,720 pages)	83.65	single run · official prompt and scoring (text, formula CDM, table TEDS) · temperature 0 · thinking on	#1

Evaluation protocol

Common to every benchmark

Setting	Value
Thinking budget	131,072 tokens (max generated tokens per sample)
Sampling	temperature 1.0 · topp 0.95 · topk 20
Precision	bf16
Engine	vLLM, tensor parallel 8 (or 4), expert parallel

Per benchmark

Benchmark	Samples per question	Reported score
AIME 2026	16	majority vote (maj@16); mean over 16 = 98.75
HMMT Feb 2026	16	majority vote (maj@16); mean over 16 = 96.59
GPQA Diamond	up to 16	majority vote
MMLU-Pro	1	single sample (no voting)
MMMU-Pro (vision)	3	majority vote (maj@3)
IFStruct	1	single run, temperature 0, 16,000 max tokens (official harness defaults)
MDPBench	1	single run, temperature 0, 32,768 max tokens, thinking on (pages with no transcription re-read with thinking off)

All numbers are self-measured and reproducible with the settings above. Majority-vote scores are system scores (several samples per question) and are labeled as such.

MMLU-Pro by category (single sample) — strongest in math 95.0 · biology 94.6 · physics 92.5; room to grow in law and history.

⚙️ Specifications


Architecture	Mixture-of-Experts, hybrid attention (36 linear-attention + 12 full-attention layers)
Layers / hidden	48 / 2,560
Experts	512 routed (10 active per token) + shared expert
Context	262,144 tokens
Vocabulary	248,320
Modalities	image + text → text
Precision	bf16 (~336 GB)

🚀 Quickstart

Serving with vLLM (8 × B200 or equivalent)

bash

vllm serve FINAL-Bench/Darwin-180B-RSI \
  --tensor-parallel-size 8 --enable-expert-parallel \
  --max-model-len 135168 --trust-remote-code

Chat Completions (OpenAI-compatible)

python

from openai import OpenAI
c = OpenAI(base_url="http://localhost:8000/v1", api_key="-")
r = c.chat.completions.create(model="FINAL-Bench/Darwin-180B-RSI",
    messages=[{"role": "user", "content": "Explain why the sky is blue in two sentences."}],
    temperature=1.0, top_p=0.95, extra_body={"top_k": 20})
print(r.choices[0].message.content)

Transformers

python

from transformers import AutoProcessor, AutoModelForImageTextToText
model_id = "FINAL-Bench/Darwin-180B-RSI"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(model_id, torch_dtype="auto", device_map="auto")

Tip: this is a thinking model. Give it room — a thinking budget of 32K–131K tokens is recommended for hard reasoning. Short budgets truncate the reasoning and cost accuracy.

⚠️ Limitations and disclosure

—Scores are self-measured with the settings stated in the Results table; majority-vote numbers use several samples per question.
—Very long reasoning is normal for hard problems; a short thinking budget will truncate answers and lower accuracy.
—Like every LLM, the model can be confidently wrong — use the ZTC confidence readout to gate high-stakes actions.

🔗 Related Darwin Models

—[Darwin-180B-RSI-R3](https://huggingface.co/FINAL-Bench/Darwin-180B-RSI-R3): the next RSI round of this model, ExtractBench #1 and EvasionBench #3
—[Darwin-397B-ZTC](https://huggingface.co/FINAL-Bench/Darwin-397B-ZTC) — 397B MoE (FP8), GPQA Diamond 93.43 %, ZTC on board
—[Darwin-28B-REASON](https://huggingface.co/FINAL-Bench/Darwin-28B-REASON) — 28B, GPQA Diamond 89.39 %
—[Darwin-36B-Opus](https://huggingface.co/FINAL-Bench/Darwin-36B-Opus) — 36B MoE, GPQA Diamond 88.4 %
—[Darwin-27B-RSI](https://huggingface.co/FINAL-Bench/Darwin-27B-RSI) — 27B, the first Darwin RSI model
—[Darwin-9B-NEG](https://huggingface.co/FINAL-Bench/Darwin-9B-NEG) — 9B with Negentropy distillation, GPQA Diamond 84.3 %
—[ZTC-Judge-27B](https://huggingface.co/FINAL-Bench/ZTC-Judge-27B) — standalone ZTC judge

📚 Citation

bibtex

@misc{darwin180b_rsi_2026,
  title  = {Darwin-180B-RSI: Recursive Self-Improvement on Verified Answers for a 180B Mixture-of-Experts Reasoning Model},
  author = {FINAL-Bench / Darwin Research Team},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/FINAL-Bench/Darwin-180B-RSI}},
  note   = {GPQA Diamond 94.44 \% · MMLU-Pro 88.12 \%}
}

@misc{darwin_family_2026,
  title  = {Darwin Family: MRI-Trust-Weighted Evolutionary Merging for Training-Free Scaling of Language-Model Reasoning},
  author = {Kim, Taebong and Hong, Youngsik and Kim, Minsik and Choi, Sunyoung and Jang, Jaewon and Shin, Junghoon},
  year   = {2026},
  eprint = {2605.14386},
  archivePrefix = {arXiv}
}

@misc{latin_square_2026,
  title  = {Placement Is Free, Composition Is Not: The Latin Square as a Provably-Balanced Construction for Heterogeneous Sequence-Mixer Stacks},
  author = {Kim, Taebong and Hong, Youngsik and Kim, Minsik and Choi, Sunyoung and Jang, Jaewon and Kim, Minseo},
  year   = {2026},
  eprint = {2609.20269},
  archivePrefix = {arXiv}
}

📜 License

Darwin-180B-RSI is a derivative of Qwen3.8-Flash-Next and is distributed under the Qwen Community License 1.0 (see LICENSE).

🏢 About

Built by [VIDRAFT](https://vidraft.net) · evaluated with FINAL-Bench.

This model is part of the Darwin Family.