DFveloper

AAIKR-3.5-Pro

Deploy Dedicated

README

License: apache-2.0

Key Features

Reasoning Focused: Rather than a simple generative model, this is a reasoning-specialized architecture that derives answers through logical chains of thought (Chain-of-Thought). It combines the subtle nuances that arise in Korean context with logical structure.
Korean-Centric Dataset: Pre-trained and fine-tuned primarily on refined Korean datasets to deeply understand the grammatical characteristics, cultural background, and specialized terminology of the Korean language.
Efficient 26B Architecture: Uses 26B parameters to maximize reasoning efficiency, optimized to achieve high-performance CoT (reasoning) results with relatively little VRAM.
Multi-task Capabilities: Proficient in a variety of advanced tasks including mathematical thinking, programming, literary reasoning, and interpretation of legal and technical documents.

Benchmarks

work in progress

Model Architecture

AAIKR 3.5 Pro is a decoder-only Transformer model with a large-scale 26B parameter layer. Through LOOP's proprietary algorithm, Korean token processing efficiency has been improved by more than 40%, maintaining reasoning consistency even over long contexts (context windows).

Parameters: 26 Billion
Context Window: 262k tokens
Focus: Korean Language Understanding, Logical Reasoning, Mathematical Solving

Training Process

AAIKR 3.5 Pro was completed through the following three stages:

Continuous Pre-training: Trained on a vast collection of Korean grammar textbooks, news articles, specialized books, and publicly available web data.
Supervised Fine-Tuning (SFT): Trained on a meticulously designed Korean reasoning tutorial dataset to form a coherent thought chain.
Reasoning Reinforcement Learning: Optimized through a Korean logic-verification loop that reflects human preferences, so the model provides logically sound explanations (rationales) rather than simply outputting answers.

Getting Started

python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "DFveloper/AAIKR-3.5-Pro"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", torch_dtype="auto")

prompt = """There is a problem. If a product with a 37% discount rate is priced at 15,000 won, what was the original price? Explain step by step, logically."""
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")

outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Usage Note

AAIKR 3.5 Pro can make mistakes. Please cross-check important information.

Model Family

AAIKR 3.5 Family

???
Mini → Edge devices
Basic → Everyday use
Omni → Balanced
Turbo → Fast inference (On Research)
Large → Advanced reasoning
Pro → Professional-grade (This model)
Ultra → Frontier model (Closed Source)

Infinite Family

???
Infinite Pro → Next-Gen (Closed Source)
???

Community & Contact

Developers: LOOP Research Team
Homepage: DFveloper.com
Report issues: Github Issue Link

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

DFveloper

Model Tree

Base

this model

Input Modalities

TextImage

Output Modalities

Text

Supported Functionality