Proposal: Add CAJAL — Local Scientific Paper Generation Model (Ranked #3 on Benchmark)

[Agnuxo1]

Proposal: Add CAJAL — Local Scientific Paper Generation Model (Ranked #3 on P2PCLAW Benchmark)

CAJAL — Local AI for End-to-End Scientific Paper Generation

CAJAL is a locally-runnable, open-weights LLM (4B and 9B variants) specialized for autonomous, structured scientific paper generation — from hypothesis to peer-reviewed output — with no API dependency or data exfiltration.

Why CAJAL belongs here: CAJAL-9B v2 recently ranked #3 on the P2PCLAW scientific paper generation benchmark, surpassing the majority of SOTA models. Only Claude Sonnet 4.6 ranked higher. This is achieved with a fully open, local model that runs on consumer hardware (8 GB VRAM / 4B, 16 GB VRAM / 9B), making high-quality scientific writing accessible without cloud costs or privacy risks.

Key Features

Feature	Description
Local & Private	Runs entirely offline via Ollama or llama.cpp; zero data leaves the machine
Structured Generation	Native section tokens ([ABSTRACT], [INTRO], [METHODS], [RESULTS], [DISCUSSION], [CONCLUSIONS], [REFERENCES]) produce IEEE/APA-ready papers
Scientific Quality Control	Built-in simulated peer-review (0–10 score) with automated revision loops
Citation Validation	Automatic reference verification against arXiv API
Multi-Language	Trilingual dataset (EN/ES/FR); model cards in 6 languages
Consumer Hardware	4B ≈ 8 GB VRAM; 9B ≈ 16 GB VRAM; GGUF + PyTorch formats
Open Weights & Data	Full weights, training dataset, and pipeline on HuggingFace

Quick Start

ollama pull cajal-9b   # or cajal-4b
# Generate a paper
cajal generate --topic "Quantum Error Correction" --output paper.md

Resources

• Paper: https://arxiv.org/pdf/2604.19792
• GitHub: https://github.com/Agnuxo1/CAJAL
• HuggingFace: https://huggingface.co/Agnuxo/CAJAL-9B-P2PCLAW
• PyPI: https://pypi.org/project/cajal-p2pclaw/
• Benchmark: https://www.p2pclaw.com/app/benchmark (CAJAL-9B v2 ranked Add MolCA and 3D-MoLM #3)
• Dataset: https://huggingface.co/datasets/Agnuxo/p2pclaw-training-dataset

Suggested Placement

• Section: "Language — General" or a new "Paper Generation" subsection
• Entry format:
  • CAJAL — Local scientific paper generation model (4B/9B) with structured output, citation validation, and peer-review simulation. Ranked Add MolCA and 3D-MoLM #3 on P2PCLAW benchmark. Paper | Code | Model

---

Thank you for maintaining this excellent resource! 🙏