Technical Summary: A Cartography of Open Collaboration in Open Source AI

Problem Statement

While the proliferation of open Large Language Models (LLMs) has fostered a vibrant ecosystem in Artificial Intelligence (AI), the specific methods of collaboration used to develop these models—both before and after their public release—have not been systematically studied. Existing research indicates that while open models are abundant, their reuse is highly concentrated (e.g., 1% of models account for 99% of downloads), and developer engagement often follows "nomadic" patterns rather than sustained collaboration. Furthermore, most open LLMs follow a "develop-then-release" approach with minimal pre-release collaboration, leaving a gap in understanding how open LLM projects are initiated, organized, and governed across their full lifecycle. This lack of systematic knowledge limits the ability to effectively foster the open source AI ecosystem, facilitate participation, and improve governance frameworks.

Methodology

The authors conducted an exploratory qualitative study to address this gap.

• Data Collection: Semi-structured interviews were conducted with 17 developers representing 14 diverse open LLM projects.
• Sampling Strategy: The sample was selected via purposeful sampling to ensure diversity in organizational contexts (grassroots initiatives, research institutes, startups, Big Tech companies) and geographic regions (North America, Europe, Africa, Asia). The initial sample was derived from Hugging Face Hub models with high activity metrics (downloads and discussions), later supplemented with known community-driven projects (e.g., EleutherAI, BigScience Workshop) to capture a fuller picture of the landscape.
• Analysis: The study employed an abductive coding approach. Transcripts were coded by multiple authors, with dual coding and peer debriefing used to mitigate researcher bias. The analysis focused on three Research Questions (RQs):
  1. Where and how does open collaboration take place across the open LLM lifecycle?
  2. What incentivizes engagement in the collaborative development of open LLMs?
  3. How are open collaborations on open LLMs coordinated and governed?

Key Contributions

The paper makes six primary contributions to the field of Software Engineering for AI:

1. Comprehensive Lifecycle Analysis: It provides one of the first qualitative studies mapping open collaboration practices across the full development and reuse lifecycle of open LLMs, from pre-training to post-release derivative development.
2. Six Artefact Domains: It identifies and characterizes six distinct but interdependent artefact domains where collaboration occurs: models, data, software, evaluation, compute, and community engagement. This challenges the view that collaboration is centered solely on the model weights.
3. Domain Properties: It demonstrates how participation access, expertise requirements, modularity, and control shape who can contribute and how collaboration unfolds within each domain.
4. Lifecycle Dynamics: It shows that collaboration is not static but evolves, shifting from concentrated, selective engagement in early stages to broader, distributed participation after model release.
5. Conceptual Model: It synthesizes findings into a conceptual model linking artefact domains, domain properties, lifecycle dynamics, motivations, and governance structures.
6. Practical Recommendations: It derives specific recommendations for diverse stakeholders (researchers, companies, policymakers, platform providers) to support the global open source AI community.

Key Results

1. Collaboration Across the Lifecycle (RQ1)

Collaboration patterns vary significantly by stage and artefact domain:

• Pre-training: Collaboration is concentrated in core domains (models, data, compute) and is often limited by proprietary considerations, regulatory constraints, and high resource/expertise requirements. Collaboration here relies on strategic partnerships (e.g., between foundation model developers and specialized regional groups) and access to public supercomputing infrastructure.
• Post-training: Collaboration remains relatively limited and internal, though some sharing of intermediate checkpoints with trusted partners and specialized datasets occurs. Evaluation is the most collaborative aspect, relying on community benchmarks.
• Post-release Reuse: Collaboration expands significantly into enabling domains (software, evaluation, community). Platforms like Hugging Face facilitate widespread distribution, derivative development (fine-tuning, quantization), and community feedback. This stage is characterized by broader, distributed participation.

2. Motivations (RQ2)

Developers are driven by a triad of motivations:

• Social: Democratizing access to AI, expanding language and cultural representation for underrepresented communities, knowledge sharing, and public accountability for publicly funded research ("public money, public AI").
• Economic: Building ecosystems to compete with leading AI companies, resource efficiency through model reuse, gaining market recognition, and career advancement.
• Technological: Promoting open science and reproducibility, standardizing development frameworks, and demonstrating the competitive capabilities of small models.

3. Governance and Coordination (RQ3)

The study identifies five organizational models governing open LLM projects:

1. Single Company-led: Centralized control (e.g., Meta's Llama, Hugging Face's SmolLM) with selective external collaboration.
2. Single Research Institute-led: Driven by one institute (e.g., AI Singapore, Ai2), often involving formal bureaucratic partnerships.
3. Multi-Organizational Research Institute-led: Structured collaborations across multiple entities (e.g., OpenGPT-X), often grant-funded.
4. Non-profit-sponsored Grassroots: Hybrid models combining central control of critical infrastructure with decentralized community contributions (e.g., EleutherAI, SpeakLeash).
5. Company-sponsored Grassroots: Company-coordinated initiatives mobilizing large volunteer communities (e.g., BigScience Workshop, Aya).

Governance structures range from centralized to decentralized, often utilizing platforms like Discord, Slack, and Hugging Face Hub for coordination.

Significance and Claims

The paper concludes that openness in open source AI is not a uniform property but an emergent outcome of how collaboration is organized across interconnected artefact domains, lifecycle stages, and institutional contexts.

• Theoretical Significance: The study challenges narrow, model-centric views of open source AI, arguing that a broader perspective is required to understand the interaction between different artefacts (data, software, compute) and stakeholders. It extends the understanding of open collaboration beyond traditional OSS repositories, highlighting the unique "cathedral" (pre-release) to "bazaar" (post-release) dynamics specific to LLMs.
• Practical Significance: By mapping the "cartography" of collaboration, the authors provide a framework for stakeholders to identify barriers (e.g., compute access, copyright restrictions) and leverage on-ramps (e.g., shared evaluation frameworks, community-driven data collection). The paper emphasizes that fostering the ecosystem requires addressing the specific constraints of different domains and lifecycle stages, rather than applying a one-size-fits-all approach to open source.

The authors maintain a modest stance, acknowledging that their sample represents successful and well-known projects and may not capture smaller or failed collaborative efforts, nor emerging models like "open labs" that were in early stages during the study. However, they assert that their findings offer a foundational, first glimpse into the complex socio-technical dynamics of open LLM development.