AXELIDEA-QUON-14B-Japanese-v01 is a 14B parameter Japanese-English bilingual large language model specialized in creative thinking and idea generation, developed by Axelidea Inc.
The model employs a novel approach that distills creative thinking patterns from a vast corpus of Japanese patent literature — a treasure trove of humanity's collective ingenuity. Built on shisa-ai/shisa-v2.1-unphi4-14b (MIT License), it excels in divergent thinking, cross-domain analogical reasoning, and structured ideation.
In a comparative evaluation of non-thinking instruction-tuned models in the 14B-class, AXELIDEA-QUON-14B-Japanese-v01 achieved 1st place on JA Leaderboard (average of 7 Japanese evaluation tasks).
Model
Developer
Parameters
JA Leaderboard (AVG)
AXELIDEA-QUON-14B
Axelidea
14B
71.99
shisa-v2.1-unphi4-14b
Shisa AI
14B
71.44
Gemma-3-12B-IT
Google
12B
63.42
Phi-4
Microsoft
14B
59.30
Sarashina2-13B
SB Intuitions
13B
56.43
All models were evaluated under identical conditions (lm_eval v0.4.12.dev0, 3-shot, bfloat16, --apply_chat_template).
1. Creativity Distillation from Vast Patent Literature
Creative thinking patterns were extracted and distilled from a vast corpus of Japanese patent literature. The logical structure inherent in patent documents — problem definition, solution approach, and demonstrated effects — proved highly effective as training data for creative problem-solving capabilities.
2. 60 Creativity Expert LoRAs
Based on Torrance's (1966) TTCT (Torrance Tests of Creative Thinking) framework, 60 expert LoRAs were individually trained across 4 categories (divergent thinking, analogical reasoning, perspective shifting, constrained creativity) × 15 subdomains. Each expert specializes in a specific creative thinking style, ensuring comprehensive coverage of diverse aspects of creativity.
A proprietary quality scoring system evaluates patent-derived training data across five dimensions: Originality, Elaboration, Feasibility, Fluency, and Flexibility. This enables efficient selection of high-quality training data from large-scale patent corpora.
4. QUBO-Based Expert Team Optimization
The problem of selecting the optimal combination of experts from 60 LoRAs was formulated as a QUBO (Quadratic Unconstrained Binary Optimization) problem. This formulation unifies expert quality, skill diversity, and category balance into a single energy function. It is also directly extensible to quantum computing (quantum annealers, gate-based quantum algorithms, CMOS Ising machines).
5. Knowledge-Preserving Fine-Tuning
By freezing FFN layers that store factual knowledge and adapting only the attention mechanism, the model acquires creative capabilities while preserving the base model's general knowledge — addressing the catastrophic forgetting problem.
Instruction-following capability may be limited: The model may not reliably follow complex or highly structured instructions, as training prioritized creative generation over strict instruction compliance.
Domain bias: Primarily trained on Japanese patent data, which may introduce biases toward technical and engineering domains.
Computational requirements: 14B parameters require significant GPU memory and computational resources for inference.
Factual accuracy: The model may generate incorrect or misleading information, especially outside its training domain.
Safety: This model is not designed for safety-critical applications.
Part of the computational work in this study was performed using the TSUBAME4.0 supercomputer at Institute of Science Tokyo. We gratefully acknowledge their support.
Third-Party Licenses
This model contains weights derived from the following MIT-licensed models:
shisa-ai/shisa-v2.1-unphi4-14b: Copyright (c) Shisa.AI. MIT License.
microsoft/phi-4: Copyright (c) Microsoft Corporation. MIT License.
Citation / 引用
bibtex
1@misc{axelidea-quon-2026,
2 title={AXELIDEA-QUON-14B-Japanese-v01: A Creativity-Specialized Japanese LLM via Multi-Expert Patent Distillation},
3 author={Axelidea Inc.},
4 year={2026},
5 url={https://huggingface.co/Axelidea/AXELIDEA-QUON-14B-Japanese-v01},
6}