AWAXIS-Think-27b

Qwen3.5 Hybrid Architecture | ~26B Params | Enhanced Thinking & Reasoning | 262K Context | BF16 | Apache 2.0

---

Model Overview

AWAXIS-Think-27b is a Korean-specialized reasoning model built on FINAL-Bench/Darwin-27B-Opus, powerfully enhanced through targeted SFT (Supervised Fine-Tuning) to achieve superior chain-of-thought reasoning, multi-step logical inference, and deep Korean language understanding.

This model inherits the evolutionary merge backbone from Darwin-27B-Opus and further strengthens reasoning capabilities through carefully curated Korean training data targeting multi-step reasoning, professional-level Korean knowledge, and metacognitive self-correction.

Key Features

• Darwin-Opus foundation — Evolutionary merge backbone from VIDRAFT's Darwin-27B-Opus
• Powerful reasoning via SFT — Dramatically enhanced chain-of-thought and multi-step reasoning through targeted fine-tuning
• K-AI optimized — Trained for Korean AI Leaderboard benchmarks (KoMMLU-Pro, CLIcK, MuSR, Com2-main)
• Thinking mode — <think> tag based step-by-step reasoning for transparent problem-solving
• 262K context — Ultra-long document processing capability
• BF16 — Memory-efficient (~48GB)
• Apache 2.0 — Free for commercial use

---

Training

Item	Details
Base Model	FINAL-Bench/Darwin-27B-Opus (via Darwin-27B-KR)
Method	LoRA SFT (rank=64, alpha=128) + Full Merge
Data	1,027 Korean SFT pairs (MuSR 428 + KoMMLU-Pro 500 + Metacognitive 99)
Focus	Reasoning enhancement, Korean domain knowledge, self-correcting inference
Epochs	2
Learning Rate	2e-5 (cosine schedule)
Effective Batch	16
Target Modules	q/k/v/o_proj, gate/up/down_proj (1.17% trainable params)
Hardware	8x NVIDIA B200 (183GB each)
Training Time	~25 minutes
Final Loss	0.66
Precision	BF16

SFT Data Composition

Source	Count	Description
MuSR (Korean)	428	Multi-step reasoning: causal, temporal, spatial, counterfactual
KoMMLU-Pro	500	Korean domain knowledge: law, economics, science, history, medicine
Metacognitive	99	Self-correcting reasoning with TICOS framework
Total	1,027	All pairs include <think> reasoning tags

---

Model Specifications

Property	Value
Architecture	Qwen3.5 (GatedDeltaNet Hybrid Attention, 64-layer)
Parameters	~26B
Hidden Size	5120
Layers	64
Context Length	262,144 tokens
Precision	BF16 (~48GB)
Vocab Size	248,320
Thinking	Supported (<think> tags)
License	Apache 2.0

---

Benchmark Performance

Benchmark	AWAXIS-Think-27b	Notes
CLIcK (200)	67.0%	Korean cultural & linguistic intelligence
KMMLU-Pro (50)	66.0%	Professional Korean knowledge
Metacognitive (10)	0.605	Claude-judged reasoning quality

---

VRAM Requirements

Setup	VRAM	Notes
BF16 (native)	~48 GB	Single H100/B200 or 2x A100
4-bit quantized	~14 GB	Single RTX 4090
8-bit quantized	~26 GB	Single A6000

---

Usage

Requirements: transformers >= 4.57.0

Transformers

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

tokenizer = AutoTokenizer.from_pretrained("Anserwise/AWAXIS-Think-27b")
model = AutoModelForCausalLM.from_pretrained(
    "Anserwise/AWAXIS-Think-27b",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [{"role": "user", "content": "대한민국 헌법재판소의 역할과 권한에 대해 설명해주세요."}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=4096, do_sample=False)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

vLLM

vllm serve Anserwise/AWAXIS-Think-27b \
    --enforce-eager \
    --max-model-len 32768 \
    --dtype bfloat16

---

Lineage

Qwen/Qwen3.5-27B
    |
    v
FINAL-Bench/Darwin-27B-Opus (evolutionary merge by VIDRAFT)
    |
    v
FINAL-Bench/Darwin-27B-KR (Korean-specialized variant)
    |
    v
Anserwise/AWAXIS-Think-27b (this model — reasoning-enhanced via SFT)

---

What makes AWAXIS-Think-27b special?

AWAXIS-Think-27b is specifically designed to excel at reasoning-intensive Korean tasks:

1. Multi-step Reasoning — Trained on complex narrative-based reasoning (murder mysteries, object tracking, constraint satisfaction)
2. Professional Knowledge — Enhanced Korean domain expertise across law, medicine, economics, science, and history
3. Metacognitive Ability — Self-correcting reasoning through TICOS (Thinking, Identifying, Correcting, Output, Summarizing) framework
4. Transparent Thinking — All reasoning steps visible through <think> tags, enabling verifiable AI responses

---

Acknowledgements

• VIDRAFT / FINAL-Bench — Darwin evolutionary merge system & base model
• Qwen Team — Qwen3.5 architecture
• Anserwise — AWAXIS model series

---

Citation

@misc{awaxis_think_27b_2026,
  title        = {AWAXIS-Think-27b: Reasoning-Enhanced Korean Language Model},
  author       = {Anserwise},
  organization = {Anserwise},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Anserwise/AWAXIS-Think-27b}}
}