Skip to content

πŸ“ Proposal: Add CAJAL as Native Research/Documentation ModelΒ #411

Description

@Agnuxo1

πŸ“ 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 β†’ 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 πŸ†

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

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!

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions