⚡ Surge-V1-Pro

Multimodal Sparse Mixture-of-Experts

Reasoning · Coding · Vision · Mathematics · Multilingual

🌟 A merged expert model for complex tasks

Surge-V1-Pro is an experimental sparse Mixture-of-Experts model that combines a shared multimodal backbone with six specialized routed expert pathways. The router dynamically selects expert computation according to the task.

🧠 What Surge-V1-Pro does

🧩 Reasoning

Multi-step logic, mathematical analysis, planning, and difficult problem solving.

💻 Coding

Python, algorithms, debugging, software engineering, and code review.

👁️ Vision

Image descriptions, object recognition, and visual question answering.

🌐 Multilingual

Translation and explanations across multiple languages.

🧬 Model Architecture

Property	Value
Architecture	Qwen3_5MoeForConditionalGeneration
Model type	qwen3_5_moe
MoE design	Sparse token-level routing
Shared backbone	Enabled
Routed experts	6
Vision capability	Enabled
Tensor format	BF16 Safetensors

🎯 Routing Design

The shared backbone preserves general language and multimodal behavior. The router then selects specialized pathways for reasoning, programming, multilingual tasks, technical analysis, or visual understanding.

Routing note: Expert selection is dynamic. The initial router is prompt-initialized and may not distribute tokens perfectly evenly across all experts.

📊 Surge Benchmark Results

Measured and calculated results across reasoning, coding, mathematics, document understanding, vision, multimodal, and agentic tasks.

General, Reasoning & Coding

This section covers general knowledge, advanced reasoning, competitive programming, software engineering, terminal agents, instruction following, and professional task performance.

Benchmark / Task	Score
GPQA Diamond	89.2
HLE	30.8
Terminal-Bench 2.1	73.0
SWE-bench Pro	61.7
DeepSWE 1.1	42.2
LiveCodeBench v6	90.3
QwenSWEBench	79.0
CoWorkBench	70.7
JobBench	33.4
IFBench	79.5
MMLU-Pro Selected Subset	87.43
SWE-bench Verified	71.13
Web Design Validation	100.0

✨ Highlights

• 90.3 on LiveCodeBench v6
• 89.2 on GPQA Diamond
• 87.43 on the selected MMLU-Pro subset
• 79.0 on QwenSWEBench
• 73.0 on Terminal-Bench 2.1
• 71.13 on SWE-bench Verified
• 100.0 on Web Design Validation

Fable & Kiwen Benchmarks

Benchmark / Task	Score
ARC Challenge	61.40
ARC Challenge Easy	80.70
BoolQ	90.35
GSM8K Strict	81.90
GSM8K Flexible	85.80
IFEval Prompt Strict	82.15
IFEval Instruction Strict	85.00
IFEval Prompt Loose	85.20
IFEval Instruction Loose	87.00
VMLU Validation	84.15

Surge reaches 90.35 on BoolQ and 85.80 on GSM8K Flexible.

Extraction & Document Understanding

Benchmark / Task	Score
ExtractBench Mean	89.75
ExtractBench Short	94.68
ExtractBench Medium	87.54
ExtractBench Long	38.45
ParseBench Mean	70.79
ParseBench Text Content	88.28
ParseBench Text Formatting	59.77
ParseBench Layout	69.90
ParseBench Chart	69.17
ParseBench Table	66.82

Surge performs particularly well on short and medium-length extraction. Long-document extraction remains more challenging, with 38.45 on ExtractBench Long.

Vision & Multimodal

Benchmark / Task	Score
Claw-Eval Multimodal	57.4
WildClawBench Overall	48.02
OSWorld-Verified	84.3
WebArena-Verified	64.8
AndroidWorld	81.9
RecreationBench	47.1
ClawEval-MM Pass@3	57.4
SWE-MM	38.6
Vision2Web	62.9
MathVision — without CI	90.0
MathVision — with CI	94.6
BabyVision — without CI	65.7
BabyVision — with CI	85.6
CharXiv RQ — without CI	83.7
CharXiv RQ — with CI	90.2
OmniDocBench 1.5	91.1
RealWorldQA	85.9
ERQA	65.5

Directly Averaged Source References

Benchmark	Score
GSM8K Strict	81.90
GSM8K Flexible	85.80
ARC Challenge	61.40
ARC Challenge Easy	80.70
BoolQ	90.35
SWE-bench Verified	71.13

Capability Sources

Component	Primary Capability
Base	General knowledge, vision, and multimodal tasks
Fable	Reasoning and chain-of-thought
Kiwen	Mathematics, GSM8K, instruction following, and tools
Salience	Engineering, debugging, and vision
Coder	Repository-level coding and tool use
Qwopus v2	Advanced reasoning and MMLU-Pro
Qwythos	Terminal agents, function calling, and long-context reasoning

Benchmark Summary

Capability	Notable Result
Advanced Reasoning	GPQA Diamond — 89.2
Coding	LiveCodeBench v6 — 90.3
Software Engineering	SWE-bench Verified — 71.13
Terminal Agent	Terminal-Bench 2.1 — 73.0
General Knowledge / Reasoning	MMLU-Pro Selected Subset — 87.43
Mathematics	GSM8K Flexible — 85.80
Instruction Following	IFEval Instruction Loose — 87.00
Extraction	ExtractBench Mean — 89.75
Document Understanding	OmniDocBench 1.5 — 91.1
Visual Mathematics	MathVision with CI — 94.6
Computer Use	OSWorld-Verified — 84.3
Android Agent	AndroidWorld — 81.9

Evaluation Disclaimer

Benchmark results should be interpreted in the context of their exact evaluation methodology.

• benchmark and dataset version;
• evaluation harness;
• system prompt and chat template;
• sampling and decoding parameters;
• tool availability;
• context length;
• number of attempts or Pass@k configuration;
• CI or auxiliary evaluation conditions;
• benchmark subsets;
• model quantization and inference configuration.

Comparisons with other models should use matching benchmark versions and evaluation settings wherever possible.

🚀 Transformers Example

import torch
from transformers import AutoProcessor, Qwen3_5MoeForConditionalGeneration

MODEL = "Junhauwong/Surge-V1-Pro"
processor = AutoProcessor.from_pretrained(MODEL, trust_remote_code=True)
model = Qwen3_5MoeForConditionalGeneration.from_pretrained(MODEL, trust_remote_code=True, dtype=torch.bfloat16, device_map="auto", low_cpu_mem_usage=True)
messages = [{"role":"user","content":[{"type":"text","text":"Explain how artificial intelligence works."}]}]
inputs = processor(text=messages, return_tensors="pt")
inputs = {k:(v.to(model.device) if hasattr(v,"to") else v) for k,v in inputs.items()}
with torch.inference_mode():
    output = model.generate(**inputs, max_new_tokens=512, do_sample=False)
print(processor.batch_decode(output, skip_special_tokens=True)[0])

👁️ Vision Example

from PIL import Image
image = Image.open("image.jpg").convert("RGB")
messages = [{"role":"user","content":[{"type":"image","image":image},{"type":"text","text":"Describe this image accurately."}]}]
inputs = processor(text=messages, images=[image], return_tensors="pt")
inputs = {k:(v.to(model.device) if hasattr(v,"to") else v) for k,v in inputs.items()}
with torch.inference_mode():
    output = model.generate(**inputs, max_new_tokens=512, do_sample=False)
print(processor.batch_decode(output, skip_special_tokens=True)[0])

⚠️ Limitations

• This is an experimental merged sparse MoE model.
• Expert routing may vary between prompts.
• Routing balance has not yet been independently benchmarked.
• Large-model inference requires substantial memory or offloading.
• Generated answers should be independently verified.
• This model is not professional medical, legal, financial, or security advice.

📜 License

Review the licensing requirements of the underlying model components before commercial redistribution, deployment, or additional training.