FINAL-Bench

Darwin-28B-Opus

README

License: apache-2.0

Abstract

Darwin-28B-Opus is the first reasoning model of the Darwin series built on the Qwen3.6 generation backbone. Produced by the Darwin V7 evolutionary breeding engine from two publicly available parents, it combines the strong bilingual reasoning of Qwen3.6-27B with Claude Opus 4-style chain-of-thought distilled behaviour.

On the GPQA Diamond graduate-level reasoning benchmark (198 PhD-level questions), Darwin-28B-Opus scores 88.89 % under the standard 3-stage adaptive evaluation, slightly edging out its larger MoE sibling Darwin-36B-Opus (88.4 %) and clearly surpassing its Qwen3.5-generation counterpart Darwin-27B-Opus (86.9 %).

🧬 Model Lineage

Table with columns: Role, Model, Role in the Merge
Role	Model	Role in the Merge
Father (父)	`Qwen/Qwen3.6-27B`	Qwen3.6 generation dense backbone with hybrid linear/full attention.
Mother (母)	`rico03/Qwen3.6-27B-Claude-Opus-Reasoning-Distilled`	Claude Opus reasoning-distilled variant of the same backbone (Jackrong-style distillation, 14 k traces).
Offspring	`Darwin-28B-Opus` (this model)	Darwin V7 evolutionary merge; Qwen3.6 architecture retained, Opus reasoning style inherited.

Why 28B? The 28B label denotes the Qwen3.6-generation member of the Darwin lineup (+1 over the Qwen3.5-era Darwin-27B-Opus). The actual parameter count is 27.6 B, and the architecture exactly follows Qwen3.6-27B.

⚙️ Technical Specifications

Table with columns: Component, Value
Component	Value
Architecture	`Qwen3_5ForConditionalGeneration` (Qwen3.6 generation, hybrid linear + full attention)
Parameters	27.6 B (BF16)
Hidden size	5 120
Intermediate size	17 408
Head dim	256
Layers	64 (3 linear : 1 full attention, `full_attention_interval = 4`)
Precision	bfloat16
Context length

🏆 Benchmark — GPQA Diamond (198 questions)

Darwin-28B-Opus is evaluated under our standard 3-stage adaptive evaluation protocol, identical to the protocol used across the Darwin series.

Table with columns: Stage, Decoding Protocol, Cost, Accuracy
Stage	Decoding Protocol	Cost	Accuracy
Stage 1	Single-shot greedy baseline	1×	74.75 % (148 / 198)
Stage 2	Majority vote ×8 at temperature 0.7 on Stage-1 wrongs	8×	83.84 % (166 / 198)
Stage 3	Adaptive ensemble refinement (close-tie tiebreaker + iterative MTI on residual hard questions)	≈ 20×	🥇 88.89 % (176 / 198)

Key performance indicators:

Stage 1 → Stage 3: +14.14 %p through adaptive protocol
vs Darwin-27B-Opus (86.9 %): +1.99 %p
vs Darwin-36B-Opus (88.4 %): +0.49 %p
vs Darwin-31B-Opus (85.9 %): +2.99 %p

🚀 Usage

Standard inference (Stage 1 baseline)

python
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

tok = AutoTokenizer.from_pretrained(
    "FINAL-Bench/Darwin-28B-Opus",
    trust_remote_code=True,
)
model = AutoModelForCausalLM.from_pretrained(
    "FINAL-Bench/Darwin-28B-Opus",
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)

messages = [
    {"role": "user",
     "content": "Solve: If f(x) = x³ − 3x + 2, find all critical points and classify them."}
]
text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tok(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=2048, do_sample=False)
print(tok.decode(outputs[0][inputs.input_ids.shape[-1]:], skip_special_tokens=True))

Enhanced accuracy (Stage 2-3 adaptive)

For leaderboard-grade accuracy, combine:

Stage 1 greedy baseline,
Stage 2 maj@8 temperature sampling on low-confidence answers,
Stage 3 adaptive refinement on still-disputed answers.

Reference implementation is provided in the Darwin-series evaluation harness.

🎯 Recommended Use-Cases

Graduate-level STEM reasoning (GPQA / science qualifying exams)
Mathematical problem solving (MATH, AIME-style problems)
Code generation and debugging (HumanEval, MBPP)
Complex multi-step chain-of-thought tasks
Bilingual reasoning (strong English + Korean; also Chinese / Japanese)

⚠️ Limitations

At 27.6 B parameters in bfloat16, full inference requires ≈ 55 GB of VRAM (e.g., a single A100-80GB or B200).
Optimised for English first, with secondary support for Korean, Chinese, and Japanese.
Deep Opus-style reasoning traces tend to be verbose — control with max_new_tokens as needed.

📚 Citation

bibtex
@misc{darwin28b_opus_2026,
  title  = {Darwin-28B-Opus: Evolutionary Merging of Qwen3.6-27B with Claude-Opus-Distilled Reasoning},
  author = {FINAL-Bench / Darwin Research Team},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/FINAL-Bench/Darwin-28B-Opus}},
  note   = {Darwin V7 · Mother-centric Ratio Interpolation merge · 88.89 % GPQA Diamond (3-stage)}
}

Darwin-36B-Opus — MoE 36B, Qwen3.6-35B-A3B × Opus distilled, GPQA 88.4 %
Darwin-31B-Opus — 31B dense, multilingual-strong reasoning, GPQA 85.9 %
Darwin-27B-Opus — 27B dense (Qwen3.5 generation), GPQA 86.9 %
Darwin-9B-NEG — 9B with Native Entropy Gating, GPQA 84.3 %
Darwin-9B-Opus — the Qwen3.5-9B Darwin member
Darwin-4B-Genesis — smallest Darwin member

Want Darwin-class reasoning without a datacenter GPU? POCKET is VIDRAFT's on-device family — a 35B model that runs on a phone and on a GPU-less PC using stock llama.cpp (no fork, no CUDA, no cloud). On a free CPU it generates ~3.4× faster than Bonsai, the most-downloaded on-device model (2M+), at matched quality (HellaSwag 61.0 % vs 60.0 %, a statistical tie).

▶️ Live CPU chat (try it now): https://huggingface.co/spaces/FINAL-Bench/POCKET-35B-CPU
📚 POCKET collection: https://huggingface.co/collections/FINAL-Bench/pocket-models-6a618ee5d23eafb7e185a5c6
📦 POCKET-35B-GGUF: https://huggingface.co/FINAL-Bench/POCKET-35B-GGUF
🇰🇷 POCKET-KR-MLX (iPhone / Mac): https://huggingface.co/FINAL-Bench/POCKET-KR-MLX
🇬🇧 POCKET-EN-GGUF: https://huggingface.co/FINAL-Bench/POCKET-EN-GGUF

This model is introduced in Darwin Family.

Darwin V7 · Qwen3.6 generation flagship · Sealed 2026-04-25 · FINAL-Bench

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

FINAL-Bench

Model Tree

Base

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality

Dedicated Endpoints

Explore FriendliAI today

Get started Talk to an engineer

README

License: apache-2.0

Abstract

🧬 Model Lineage

Table with columns: Role, Model, Role in the Merge
Role	Model	Role in the Merge
Father (父)	`Qwen/Qwen3.6-27B`	Qwen3.6 generation dense backbone with hybrid linear/full attention.
Mother (母)	`rico03/Qwen3.6-27B-Claude-Opus-Reasoning-Distilled`	Claude Opus reasoning-distilled variant of the same backbone (Jackrong-style distillation, 14 k traces).
Offspring	`Darwin-28B-Opus` (this model)	Darwin V7 evolutionary merge; Qwen3.6 architecture retained, Opus reasoning style inherited.

Why 28B? The 28B label denotes the Qwen3.6-generation member of the Darwin lineup (+1 over the Qwen3.5-era Darwin-27B-Opus). The actual parameter count is 27.6 B, and the architecture exactly follows Qwen3.6-27B.

⚙️ Technical Specifications

Table with columns: Component, Value
Component	Value
Architecture	`Qwen3_5ForConditionalGeneration` (Qwen3.6 generation, hybrid linear + full attention)
Parameters	27.6 B (BF16)
Hidden size	5 120
Intermediate size	17 408
Head dim	256
Layers	64 (3 linear : 1 full attention, `full_attention_interval = 4`)
Precision	bfloat16
Context length

🏆 Benchmark — GPQA Diamond (198 questions)

Darwin-28B-Opus is evaluated under our standard 3-stage adaptive evaluation protocol, identical to the protocol used across the Darwin series.

Table with columns: Stage, Decoding Protocol, Cost, Accuracy
Stage	Decoding Protocol	Cost	Accuracy
Stage 1	Single-shot greedy baseline	1×	74.75 % (148 / 198)
Stage 2	Majority vote ×8 at temperature 0.7 on Stage-1 wrongs	8×	83.84 % (166 / 198)
Stage 3	Adaptive ensemble refinement (close-tie tiebreaker + iterative MTI on residual hard questions)	≈ 20×	🥇 88.89 % (176 / 198)

Key performance indicators:

Stage 1 → Stage 3: +14.14 %p through adaptive protocol
vs Darwin-27B-Opus (86.9 %): +1.99 %p
vs Darwin-36B-Opus (88.4 %): +0.49 %p
vs Darwin-31B-Opus (85.9 %): +2.99 %p

🚀 Usage

Standard inference (Stage 1 baseline)

python
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

tok = AutoTokenizer.from_pretrained(
    "FINAL-Bench/Darwin-28B-Opus",
    trust_remote_code=True,
)
model = AutoModelForCausalLM.from_pretrained(
    "FINAL-Bench/Darwin-28B-Opus",
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)

messages = [
    {"role": "user",
     "content": "Solve: If f(x) = x³ − 3x + 2, find all critical points and classify them."}
]
text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tok(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=2048, do_sample=False)
print(tok.decode(outputs[0][inputs.input_ids.shape[-1]:], skip_special_tokens=True))

Enhanced accuracy (Stage 2-3 adaptive)

For leaderboard-grade accuracy, combine:

Stage 1 greedy baseline,
Stage 2 maj@8 temperature sampling on low-confidence answers,
Stage 3 adaptive refinement on still-disputed answers.

Reference implementation is provided in the Darwin-series evaluation harness.

🎯 Recommended Use-Cases

Graduate-level STEM reasoning (GPQA / science qualifying exams)
Mathematical problem solving (MATH, AIME-style problems)
Code generation and debugging (HumanEval, MBPP)
Complex multi-step chain-of-thought tasks
Bilingual reasoning (strong English + Korean; also Chinese / Japanese)

⚠️ Limitations

At 27.6 B parameters in bfloat16, full inference requires ≈ 55 GB of VRAM (e.g., a single A100-80GB or B200).
Optimised for English first, with secondary support for Korean, Chinese, and Japanese.
Deep Opus-style reasoning traces tend to be verbose — control with max_new_tokens as needed.

📚 Citation

bibtex
@misc{darwin28b_opus_2026,
  title  = {Darwin-28B-Opus: Evolutionary Merging of Qwen3.6-27B with Claude-Opus-Distilled Reasoning},
  author = {FINAL-Bench / Darwin Research Team},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/FINAL-Bench/Darwin-28B-Opus}},
  note   = {Darwin V7 · Mother-centric Ratio Interpolation merge · 88.89 % GPQA Diamond (3-stage)}
}

Darwin-36B-Opus — MoE 36B, Qwen3.6-35B-A3B × Opus distilled, GPQA 88.4 %
Darwin-31B-Opus — 31B dense, multilingual-strong reasoning, GPQA 85.9 %
Darwin-27B-Opus — 27B dense (Qwen3.5 generation), GPQA 86.9 %
Darwin-9B-NEG — 9B with Native Entropy Gating, GPQA 84.3 %
Darwin-9B-Opus — the Qwen3.5-9B Darwin member
Darwin-4B-Genesis — smallest Darwin member

▶️ Live CPU chat (try it now): https://huggingface.co/spaces/FINAL-Bench/POCKET-35B-CPU
📚 POCKET collection: https://huggingface.co/collections/FINAL-Bench/pocket-models-6a618ee5d23eafb7e185a5c6
📦 POCKET-35B-GGUF: https://huggingface.co/FINAL-Bench/POCKET-35B-GGUF
🇰🇷 POCKET-KR-MLX (iPhone / Mac): https://huggingface.co/FINAL-Bench/POCKET-KR-MLX
🇬🇧 POCKET-EN-GGUF: https://huggingface.co/FINAL-Bench/POCKET-EN-GGUF

This model is introduced in Darwin Family.

Darwin V7 · Qwen3.6 generation flagship · Sealed 2026-04-25 · FINAL-Bench

Darwin-28B-Opus

README

Abstract

🧬 Model Lineage

⚙️ Technical Specifications

🏆 Benchmark — GPQA Diamond (198 questions)

🚀 Usage

Standard inference (Stage 1 baseline)

Enhanced accuracy (Stage 2-3 adaptive)

🎯 Recommended Use-Cases

⚠️ Limitations

📚 Citation

🔗 Related Darwin Models

📱 Related — POCKET (run a 35B model on-device)

Explore FriendliAI today

README

Abstract

🧬 Model Lineage

⚙️ Technical Specifications

🏆 Benchmark — GPQA Diamond (198 questions)

🚀 Usage

Standard inference (Stage 1 baseline)

Enhanced accuracy (Stage 2-3 adaptive)

🎯 Recommended Use-Cases

⚠️ Limitations

📚 Citation

🔗 Related Darwin Models

📱 Related — POCKET (run a 35B model on-device)