Cianidos/Darwin-180B-RSI-R3-Sushi-2.5bpw

Darwin-180B-RSI-R3 — Sushi K2.5 / affine8

A Sushi pack converted directly from the original Darwin R3 BF16 weights, for local inference on Apple Silicon. Tested on an Apple M5 Pro with 48 GiB unified memory, using SSD expert streaming.

This is a Sushi-format pack, not a native ExLlamaV3 checkpoint, GGUF, or ordinary Transformers checkpoint. Use Sushi 1.3.0; other versions have not been tested here. The n-gram table, vision weights and native MTP head are included. No additional model files are required.

Everything above the Original model card divider describes this conversion. The preserved upstream card below describes the original Darwin models and their publishers' results, not measured quality of this quantization.

Pack and conversion parameters

• Source: FINAL-Bench/Darwin-180B-RSI-R3, pinned revision 15495159641ac68386ccf446e34e01fb1ad86b53.
• Converter/runtime: Sushi v1.3.0, release commit 8e4c0af1f5d239dcbed3d58728259432c86f1d96; MLX 0.32.3, NAX enabled.
• Hardware: Apple M5 Pro, 48 GiB unified memory, macOS 27.0.1, internal SSD, AC power.
• Routed experts: EXL3 K2.5, MCG codebook, window 15, including the MTP expert bank. All 512 experts per layer retained; top-10 routing unchanged. No expert pruning or additional fine-tuning.
• Dense trunk: MLX affine 8-bit, group size 64, with floating tensors retained as the converter requires.
• N-gram table: affine 4-bit, group size 32, in ngram_table.bin.
• Calibration: Sushi's default 65,536-token prose/code mix; 16 experts per batch; float64 Hessian factorization. Calibration-token SHA-256: 7f2a5b1413545d9daf73ef13f06769914c68693121cd86f1e2611bf8d5c419c3.
• Native context: 262,144 tokens, prompt plus generated response. Native RoPE preserved; no 1M-context extension.
• Conversion time: about 3 h 54 min on this Mac. The process reported 16.88 GB peak physical footprint during conversion; this is not an inference-memory requirement.

2.5bpw names the routed-expert code rate, not an average over every tensor or total download size. This is a new quantization of Darwin R3, not a repack of scorpoon's EXL3 SC build and not a reproduction of Sushi's privately tuned official pack recipes.

Size on disk

Component	Decimal GB	GiB
Indexed weight shards (52 files)	45.33	42.22
Disk-backed n-gram table	32.00	29.80
Complete download, approximately	77.36	72.05

The original BF16 checkpoint is not needed to run this pack. Download every shard and ngram_table.bin; do not omit the table because it is not a safetensors shard.

Install and download

Apple Silicon and macOS 26.2 or newer are required by this Sushi build.

brew install beamivalice/tap/sushi
sushi --version

hf download Cianidos/Darwin-180B-RSI-R3-Sushi-2.5bpw \
  --local-dir "$HOME/.sushi/models/Darwin-180B-RSI-R3-Sushi-2.5bpw"

Use hf download for this community pack: Sushi 1.3.0's built-in sushi pull accepts only the beamster namespace.

Run on a 48 GiB Mac

Measured text configuration

This is the configuration used for the performance measurements below: 32K configured context, 22 GiB streaming budget, KV8, native MTP, no vision and no prefix-cache reuse.

sushi serve \
  --model "$HOME/.sushi/models/Darwin-180B-RSI-R3-Sushi-2.5bpw" \
  --host 127.0.0.1 --port 12345 \
  --ssd-budget-gb 22 --ctx-size 32768 \
  --prefill-chunk 1024 --kv-quant 8 \
  --mtp --mtp-head-kv-quant \
  --no-vision --no-drafter \
  --expert-pick-tolerance 0 \
  --prefix-cache-entries 0 --metrics

Open http://127.0.0.1:12345/ for Sushi's browser chat and dashboard. OpenAI-compatible API: http://127.0.0.1:12345/v1; Anthropic-compatible API: http://127.0.0.1:12345.

Maximum native context, vision and SSD prefix reuse

This second configuration passed startup and advertised the full context with vision loaded on the same 48 GiB Mac. The owner also reports that it works in use. A full 262K-token prompt combined with images has not been stress-tested, and these are not the settings of the benchmark below.

On this Mac the owner explicitly set the GPU wired-memory allowance to 40 GiB. This is a system-wide setting, not a reservation or a per-process RAM cap. It resets at reboot. Raising it can put other applications and macOS under memory pressure; do not copy the value onto a smaller Mac.

# Only for a 48 GiB Mac with sufficient memory available:
sudo sysctl iogpu.wired_limit_mb=40960

sushi serve \
  --model "$HOME/.sushi/models/Darwin-180B-RSI-R3-Sushi-2.5bpw" \
  --host 127.0.0.1 --port 12345 \
  --ssd-budget-gb 28 --wired-margin-gib 2 \
  --ctx-size 262144 --prefill-chunk 1024 \
  --kv-quant 8 --mtp --mtp-head-kv-quant \
  --vision --no-drafter \
  --expert-pick-tolerance 0 \
  --prefix-cache-entries 4 --prefix-cache-mem 0 \
  --prefix-cache-disk 100GB --metrics

Restore the macOS default allowance with sudo sysctl iogpu.wired_limit_mb=0.

What the memory flags mean

• --ssd-budget-gb is a RAM residency target, in GiB, for the resident trunk, enabled MTP/vision, streaming buffers and expert cache. It is not SSD storage space and not a hard cap on total process memory. Active KV/recurrent state, temporary computation, CPU allocations and the OS file cache need additional memory.
• At the measured 22 GiB, text-only/MTP setup, the startup ledger allocated 16.05 GB for experts: 215 of 512 cache slots per layer. Missing experts are read from SSD; none are permanently removed.
• At 28 GiB with vision, the startup ledger allocated 21.57 GB for experts: 289 of 512 slots per layer. The 2 GiB wired margin leaves a 38 GiB planning ceiling under the owner's 40 GiB allowance; it does not guarantee that total process footprint stays below 38 GiB.
• --expert-pick-tolerance 0 preserves exact expert routing. There is no lossy substitution of a cached expert for a selected expert.
• --prefix-cache-disk 100GB is a 100 GiB disk-capacity limit, not an immediate 100 GiB allocation. --prefix-cache-mem 0 disables idle RAM retention; matching prompt state is restored from SSD into RAM for the active request. Prefix state can survive a server restart. This is separate from the model-weight expert cache.
• A larger context, images, other loaded models or other applications may require lowering the streaming budget. Memory refusals should be addressed by reducing the plan, not disabling admission checks.

For byte-stable comparisons, keep prefix caching off: hybrid Qwen cache restoration can change rounding relative to a cold prefill. Ordinary chat benefits from prefix reuse.

Performance measured on the M5 Pro / 48 GiB

Date: 2026-10-25. Sushi 1.3.0, release binary SHA-256 6e2527f18c8f214976cdb9a979eb919196599686e3678595eeab93c4d28441e3.

Measured with llmprobe 0.6.17, greedy, reasoning off, three measured runs after warmup, using the text configuration above. Processes used foreground QoS (taskpolicy -a); AC power, no concurrent model server, a three-minute initial cooldown, system automatic fans. No manual fan control or die-temperature measurement. Serial and MTP ran sequentially in the same session, not as an interleaved A/B.

Mode	Decode tok/s, median (min–max)	Prefill tok/s, median (min–max)
Plain decoding (--no-mtp)	26.2 (24.9–26.5)	412.9 (411.7–413.4)
Native MTP (--mtp --mtp-head-kv-quant)	36.0 (35.3–38.0)	399.8 (392.6–400.0)

Headline prefill workload: 10,406 input tokens. Separate context-ladder results:

Actual input tokens	Serial prefill / decode tok/s	MTP prefill / decode tok/s
512	171 / 22.6	170 / 31.1
2,109	288 / 21.2	283 / 29.3
8,250	279 / 21.2	274 / 29.7

At 8,250 input tokens, the MTP arm's median time to first token was 30.1 seconds. The harness classified sustained-load drift as steady in both arms. It also reported that the engine ignored ignore_eos/min_tokens, so decode samples followed the model's own stopping behavior and did not all generate equal-length outputs. These custom runs are not comparable to default llmprobe runs or anecdotal Strata results.

The MTP arm's MLX allocator peak was 24.38 GiB; this is not total process or system memory. The full-context/vision configuration above was only startup-checked separately and has no matching performance result here.

Benchmark invocation, once per server mode:

npx --yes llmprobe@0.6.17 http://127.0.0.1:12345 \
  -m Darwin-180B-RSI-R3-Sushi-2.5bpw \
  --bench-only --reasoning off --rungs 512,2k,8k --runs 3 \
  --no-color --no-save --save bench.json

The measured local pack directory was named Darwin-180B-RSI-R3-Sushi-K2.5; the download command above uses this repository's name, so the model id in the example follows that directory instead. Weight bytes are unchanged.

Validation and quality limits

• Pack structure checked: 3,069 indexed tensors, valid safetensors spans, 52 indexed shards, all 48 routed-expert layers and MTP present.
• 5/5 serial and 6/6 MTP functional checks passed: chat, stream/non-stream bytes on one prompt, a forced tool call with typed arguments, JSON output, Anthropic Messages and one MTP/serial greedy answer comparison.
• Five one-shot reasoning/debugging questions gave correct final answers. One scheduling answer had inconsistent table intervals despite the correct optimum and critical-path proof. This is a smoke test, not a standardized model-quality benchmark.
• No BF16 KLD or perplexity comparison was performed. The original Darwin benchmark scores below have not been reproduced for this quantization. Quantization is lossy; no BF16-equivalence claim is made.
• Vision weights loaded successfully in the second configuration; no controlled image-quality evaluation was performed here. Native context is 262K, but the longest timed prompt in this report was 10,406 tokens.
• No external DFlash2 drafter was used or tested. The native MTP head is included and was observed engaging with exact acceptance.
• No ZTC confidence probe or extra ZTC-serving integration is included; those claims belong to the original model card.

Conversion command

The following parameters were used with the released Sushi 1.3.0 binary. Replace the paths with your local source and output directories. The original BF16 source is about 360 GB; it does not have to fit in RAM because conversion is layer-by-layer. Leave disk room for the source, output and scratch files. Conversion supports --resume.

hf download FINAL-Bench/Darwin-180B-RSI-R3 \
  --revision 15495159641ac68386ccf446e34e01fb1ad86b53 \
  --local-dir /path/to/Darwin-R3-BF16

caffeinate -i /usr/bin/time -l sushi convert exl3 \
  --model /path/to/Darwin-R3-BF16 \
  --out /path/to/Darwin-R3-Sushi-K2.5 \
  --experts-k 2.5 --codebook mcg --window 15 \
  --trunk-bits 8 --trunk-group 64 \
  --calib-tokens 65536 --batch 16 --factor-precision 64

Credits and license

Darwin training and upstream evaluations: FINAL-Bench / VIDRAFT. Parent architecture and model: Qwen. Converter, runtime and pack format: Sushi / beamivalice. The original Qwen Community License 1.0 is included unchanged in LICENSE; its commercial-use conditions apply. No broader license is granted by this conversion.

---

Original model card

The following body is preserved from the pinned BF16 source README. Its BF16/vLLM/Transformers instructions describe the original model, not this Sushi pack. Benchmark claims and comparisons are the original publishers' statements.

Darwin-180B-RSI-R3

180B Mixture-of-Experts · vision-language · #1 on ExtractBench (90.29) · the Darwin-180B-RSI line now holds ten Hugging Face official #1s: nine by Darwin-180B-RSI (R1) and ExtractBench by R3 · self-improving

🥇 R3 is #1 on the ExtractBench leaderboard (90.29), ahead of its own parent Qwen3.8-Flash-Next (89.88), and #3 on EvasionBench (77.83).

💻 Run it on your own machine: 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 is identical to BF16 (87.65%).

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

The second round of model-level self-improvement on top of Darwin-180B-RSI. R3 continues training from R1 with the same recipe: the model solves verifiable problems, keeps only its own solutions that check out as correct, and trains on them. No human-written solutions or reasoning traces are used. R3 is released so anyone can download it, run it, and check the numbers below.

---

🏆 Ten #1s in the Darwin-180B-RSI line

Hugging Face official benchmark leaderboards. Each score is listed under the model that produced it.

Benchmark	Score	Model	Leaderboard
ExtractBench (370 documents)	90.29	R3 (this model)	#1
GPQA Diamond (198)	94.44	R1 (Darwin-180B-RSI)	#1
MMLU-Pro (12,032)	88.12	R1	#1
AIME 2026 (30)	100.0	R1	#1
HMMT Feb 2026 (33)	100.0	R1	#1
MMMU-Pro (vision, 1,730)	79.48	R1	#1
LEXam (law, MCQ 4-choice, 1,655)	68.94	R1	#1
LEXam-hard (law, open-ended, 518)	45.72	R1	#1
IFStruct (structured output, 2,000 prompts)	98.95	R1	#1
MDPBench (multilingual document parsing, 2,720 pages)	83.65	R1	#1

R3's own leaderboard entries:

Benchmark	R3	Leaderboard	Setting
ExtractBench (370 documents)	90.29, #1	llamaindex/ExtractBench	official harness, 32,768 max tokens (same as the parent), temperature 0, thinking off, single run
EvasionBench (16,726 questions)	77.83, #3	FutureMa/EvasionBench	inspect-ai task from the dataset eval.yaml, temperature 1.0, top_p 0.95, 8,192 max tokens, thinking on, single run

Full settings are recorded in .eval_results/.

R1 head-to-head with Chinese frontier models

Model	AIME 2026	GPQA Diamond	MMLU-Pro	MMMU-Pro	HMMT Feb 2026	LEXam	LEXam-hard
🧬 Darwin-180B-RSI, R1 (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	·

_{Scores as listed on the Hugging Face official benchmark leaderboards (self-reported by each model's publisher). "·" = not reported. Open-weight models only; closed API models are not included. Settings (samples, voting, thinking budget) differ across models. R1 settings are in the evaluation protocol below.}

---

🧬 What changed from R1

• Starting point: R1 weights (not the parent). R3 is a true second round.
• Practice problems: 3,000 SuperGPQA questions (middle and hard difficulty) that were never used in R1 training. R1 solved each one 8 times.
• What it learned from: only the "boundary" problems, where R1 was right on 2 to 6 of 8 attempts (462 problems). From those, up to 2 of R1's own correct, untruncated solutions per problem (714 solutions in total).
• What was trained: the same components as R1 (attention paths and shared experts) via LoRA, then merged. All 512 routed experts, the router and the vision encoder are unchanged.
• No benchmark data: GPQA Diamond and the held-out set below were never used for training or selection.

Held-out SuperGPQA, 1,000 questions never used in training or selection

4 samples per question, 16K thinking budget, temperature 1.0.

Model	Single sample	Mean of 4	Majority of 4
R1 (Darwin-180B-RSI)	65.30	65.67	68.30
R3 (this model)	66.30	66.70	69.00

Paired per-question difference, R1 → R3 (mean of 4): +1.03 points, 95% CI [+0.05, +2.00]. The second round of self-improvement produced a measurable gain over R1.

GPQA Diamond, 198 questions

8 samples per question, 32K thinking budget, temperature 1.0.

Model	Single sample	Mean of 8	Majority of 8
R0 (parent, Qwen3.8-Flash-Next)	84.85	85.35	90.91
R1 (Darwin-180B-RSI)	84.85	85.80	89.90
R3 (this model)	85.86	86.05	90.40

---

🧬 The Darwin Family

Darwin is VIDRAFT's measurement-driven reasoning model family: 50+ official models, 400+ community derivatives, and 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; the RSI line adds a new ingredient: the model's own verified work.
• Measured, not claimed. Every change must beat its predecessor on held-out tests before it ships.

Lineage

Role
R0, parent	Qwen/Qwen3.8-Flash-Next	180B MoE vision-language backbone · Qwen Community License 1.0
R1	Darwin-180B-RSI	first RSI round: the parent's own verified solutions fed back as training signal · nine #1s
R3 (this model)	Darwin-180B-RSI-R3	second RSI round, trained from R1 on R1's own verified solutions · ExtractBench #1
Preserved	512 routed experts · router · vision encoder	untouched in every round

---

🔁 RSI: a model that improves from its own work

Recursive self-improvement (RSI) is the core of this line. 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. R1 was round one; R3 is the next round.

Practice sets are deduplicated against every evaluation set we report (8-gram overlap filter).

Model-level RSI vs. harness-level RSI

Darwin-180B-RSI-R3 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 is 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, with no extra tokens and no second model.

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

The ZTC probe published with Darwin-180B-RSI was fitted on R1's hidden states. A probe fitted on R3 will be added to this repository; until then, use R1's probe only as a rough signal.

---

📐 Evaluation protocol

R3 entries (ExtractBench, EvasionBench): see the settings column in the table above and .eval_results/. ExtractBench was run with thinking turned off (chat_template_kwargs: {"enable_thinking": false}), which suits schema-guided extraction.

R1 entries (its #1s):

Setting	Value
Thinking budget	131,072 tokens (32,768 for LEXam and LEXam-hard)
Sampling	temperature 1.0 · top_p 0.95 · top_k 20
Precision	bf16
Engine	vLLM, tensor parallel 8 (or 4), expert parallel

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)
LEXam	4	majority vote (single sample 60.54 · mean 61.42)
LEXam-hard	1	single sample, judged by DeepSeek-R1-0528 per the official eval.yaml
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.

---

⚙️ 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)

vllm serve FINAL-Bench/Darwin-180B-RSI-R3 \
  --tensor-parallel-size 8 --enable-expert-parallel \
  --max-model-len 139264 --trust-remote-code

Chat Completions (OpenAI-compatible)

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-R3",
    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)

For structured extraction (JSON to a schema), turn thinking off, as in the ExtractBench run:

r = c.chat.completions.create(model="FINAL-Bench/Darwin-180B-RSI-R3", messages=msgs, temperature=0,
    response_format={"type": "json_object"},
    extra_body={"chat_template_kwargs": {"enable_thinking": False}})

Transformers

from transformers import AutoProcessor, AutoModelForImageTextToText
model_id = "FINAL-Bench/Darwin-180B-RSI-R3"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(model_id, torch_dtype="auto", device_map="auto")

Tip: for hard reasoning, keep thinking on and give it room (a thinking budget of 32K to 131K tokens). For document extraction, thinking off was faster and scored higher in our ExtractBench runs.

---

⚠️ Limitations and disclosure

• Scores are self-measured with the settings stated above; 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 a confidence readout such as ZTC to gate high-stakes actions.

---

🔗 Related Darwin Models

• Darwin-180B-RSI: R1, the model R3 was trained from, #1 on nine Hugging Face official leaderboards
• POCKET-Darwin-180B-GGUF: 4-bit GGUF of R3 for laptops, CPU-only machines and DGX Spark
• Darwin-397B-ZTC: 397B MoE (FP8), GPQA Diamond 93.43 %, ZTC on board
• Darwin-28B-REASON: 28B, GPQA Diamond 89.39 %
• Darwin-27B-RSI: 27B, the first Darwin RSI model
• ZTC-Judge-27B: standalone ZTC judge

---

📚 Citation

@misc{darwin180b_rsi_r3_2026,
  title  = {Darwin-180B-RSI-R3: A Second Round of Model-Level Self-Improvement 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-R3}},
  note   = {ExtractBench 90.29}
}

@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}
}

---

📜 License

Darwin-180B-RSI-R3 is a derivative of Qwen3.8-Flash-Next (through Darwin-180B-RSI) and is distributed under the Qwen Community License 1.0 (see LICENSE).

🏢 About

Built by VIDRAFT · evaluated with FINAL-Bench. Part of the Darwin Family.