📝 Proposal: Add CAJAL as Native Research/Documentation Model

[Agnuxo1]

📝 Proposal: Add CAJAL as Native Research/Documentation Model for Trae Agent

Summary

CAJAL is an open-source scientific paper generation model (4B–9B parameters) that runs entirely locally via Ollama. It ranks #3 globally on scientific writing benchmarks, beating GPT-4 and Gemini.

Why Trae Agent + CAJAL?

Trae Agent already supports Ollama (great!). But CAJAL brings a new use case: instead of just coding, Trae Agent could generate:

• Research documentation from code
• README academic papers
• Methodology sections for ML repos
• Grant proposals and tech reports

Proposal

Add CAJAL as a preset model option alongside existing Ollama models:

# .trae-agent-config.yaml
provider: ollama
model: cajal-9b-v2   # Auto-downloads from Ollama library
mode: research       # New mode: code → documentation

Research Mode Workflow:

1. Agent reads codebase structure and docstrings
2. Feeds context to CAJAL via Ollama API
3. Generates: Abstract, Methodology, Results, Discussion
4. Outputs LaTeX/Markdown directly to docs/ folder

Model Availability

Model	Ollama Pull Command	Size	VRAM
CAJAL-9B-v2-Q4_K_M	ollama pull cajal-9b-v2-q4	4.7 GB	6 GB
CAJAL-9B-v2-Q6_K	ollama pull cajal-9b-v2-q6	6.8 GB	8 GB
CAJAL-9B-v2-Q8_0	ollama pull cajal-9b-v2-q8	8.9 GB	10 GB

Note: We can provide official Ollama model manifests if needed.

Benchmark Results 🏆

CAJAL-9B v2 ranks #3 on P2PCLAW Scientific Writing Benchmark:

Rank	Model	Score
1	Claude Sonnet 4.6	8.9
2	[Undisclosed]	8.7
3	CAJAL-9B v2	8.4
4	GPT-4 Turbo	8.1
5	Gemini 1.5 Pro	7.9

A 9B local model beating 70B+ cloud models on scientific writing.

Links

• Code: https://github.com/Agnuxo1/CAJAL
• Models: https://huggingface.co/Agnuxo1
• Benchmark: https://www.p2pclaw.com/app/benchmark
• License: Apache 2.0

Team

• Francisco Angulo de Lafuente (Lead, Spain)
• Vladimir Veselov (MIET Moscow)
• Seid Mehammed Abdu (Woldia University)
• Nirmal Tej Kumar (UT Dallas)

Next Steps

We're ready to:

1. Provide official Ollama manifests for all CAJAL variants
2. Submit a PR adding CAJAL preset to Trae Agent
3. Create a demo video showing Trae Agent + CAJAL generating a paper from a codebase

Would love to hear your thoughts!