BigCode Project
BigCode is an open scientific collaboration working on the responsible development and use of large language models for code. It aims to empower machine-learning and open-source communities through open governance, open datasets, open experiments, and openly released code models.
At a Glance
- Open-source developers and professional software engineers
- Citizen developers and organizations building software or workflows
- Machine-learning researchers and academic laboratories
- AI companies and enterprise engineering teams needing code models
- +1 more
AI Tools by BigCode Project
(1)
BigCode Evaluation Harness
Code Model Evaluation Framework
Discussions
No discussions yet
Be the first to start a discussion about BigCode Project
Latest News
BigCodeArena: Judging code generations end to end with code executions
BigCodeBench benchmark presented as ICLR'25 work for code generation with diverse function calls and complex instructions
StarCoder2 and The Stack v2 released: a 3B/7B/15B model family and expanded open code dataset
ServiceNow and Hugging Face released StarCoder, an open-access 15B code-generation LLM
Products & Services
Family of 3B, 7B, and 15B open code LLMs. The models use grouped-query attention, a 16,384-token context window with 4,096-token sliding-window attention, fill-in-the-middle training, and The Stack v2 data; the 15B model covers 600+ languages and more than 4T tokens.
Approximately 15B-parameter multilingual code models trained on permissively licensed GitHub code, commits, issues, and notebooks across 80+ languages. Capabilities include code completion, infilling, instruction-based code modification, natural-language code explanation, text-to-code, and technical-assistant interaction.
A 1.1B-parameter multilingual code model trained on Java, JavaScript, and Python code from The Stack, supporting left-to-right generation and infilling.
Large open code-pretraining datasets with license and language detection, repository context, opt-out support, and governance tools; The Stack v2 is derived from the Software Heritage archive.
Market Position
BigCode differentiates itself as an open-science, open-governance alternative to proprietary or more closed code-assistant/model-development efforts such as OpenAI Codex/GitHub Copilot and Amazon CodeWhisperer. Its positioning combines openly accessible multilingual code models and datasets with permissive/responsible licensing, opt-out and attribution tooling, reproducible evaluation, and community participation; its official StarCoder announcement says the models match or surpass code-cushman-001 on several benchmarks.
Leadership
Founders
Harm de Vries
ServiceNow Research researcher and lead of the Large Language Model Lab; he initiated the BigCode idea in Utrecht and served as a steering-committee co-lead.
Thomas Wolf
Hugging Face co-founder and chief science officer; he was the Hugging Face counterpart approached by Harm de Vries to establish the jointly led collaboration.
Executive Team
Harm de Vries
Steering committee co-lead
ServiceNow Research; lead of the Large Language Model Lab and initiator of the BigCode collaboration.
Leandro von Werra
Steering committee co-lead
Hugging Face machine-learning engineer and BigCode co-lead.
Founding Story
The project idea arose in Utrecht when Harm de Vries of ServiceNow Research discussed with Thomas Wolf of Hugging Face the shared vision of responsibly developing open large language models for code. Inspired by BigScience, Hugging Face and ServiceNow established a research-collaboration framework to rally the global machine-learning and open-source communities around training, studying, and releasing foundation models for code.
Business Model
Revenue Model
BigCode is an open research collaboration rather than a disclosed commercial company. Its outputs are released openly and the collaboration is resourced by Hugging Face and ServiceNow research, engineering, ethics, governance, legal, and compute resources; no paid BigCode subscription or usage-based revenue model was identified.
Target Markets
- Open-source developers and professional software engineers
- Citizen developers and organizations building software or workflows
- Machine-learning researchers and academic laboratories
- AI companies and enterprise engineering teams needing code models
- Data-governance, legal, ethics, and policy researchers working on code AI
- Professional software-engineer code completion and pair programming
- Citizen-developer application building and workflow automation
- Text-to-code generation and instruction-based code modification
- Natural-language explanation and technical assistance for code
- Research and benchmarking of code-generation language models
- Responsible dataset and model-governance research for open-source code AI