TabbyML: Open Source Competitor to GitHub Copilot Secures $3.2 Million in Funding

The competition to develop AI assistants for coding is intensifying. TabbyML, founded by two former Google employees, has raised $3.2 million in seed funding to advance its open-source code generation platform.

Unlike GitHub’s Copilot, TabbyML offers a self-hosted solution that is highly customizable. Co-founder Meng Zhang emphasizes this advantage, stating, “We envision a future where companies will demand software customization.” He shared these insights in a recent interview.

Zhang acknowledged that while proprietary software products may be more polished, open-source alternatives, like TabbyML, surpass GitHub’s OpenAI-powered tool in terms of flexibility. He noted, “Proprietary options have their limitations when compared to open-source solutions.”

Co-founder Lucy Gao highlighted that open-source software is particularly beneficial for larger enterprises. While individual developers may use open-source code, enterprise engineers typically work with proprietary code that is inaccessible to Copilot. “For instance, if a colleague writes a line of code, I can reference it instantly using TabbyML,” Gao stated.

Like many AI-driven tools, code generators can struggle with bugs or inaccuracies. Gao believes this challenge is manageable for self-hosted platforms. Each time users choose not to accept TabbyML’s suggestions or adjust its auto-filled code, the AI model improves through that feedback.

The goal of these code generators is to assist programmers rather than replace them, with positive results noted in the industry. A June survey from GitHub revealed that Copilot users accepted about 30% of the assistant’s suggestions. Zhang pointed to a more telling statistic: at a recent developer event, Google reported that 24% of its software engineers experienced over five "assistive moments" daily using its AI-enhanced internal tool, Cider.

While some decision-makers might consider reducing engineering teams after adopting a code generator, Zhang points out, “It’s a misconception. Coding isn’t a production line.”

Since its launch in April, TabbyML has garnered around 11,000 stars on GitHub and attracted investment from Yunqi Partners and ZooCap.

When discussing competition with Copilot, Zhang predicts that OpenAI's current edge will lessen as other AI models gain capability and computing costs decline.

Zhang attributes GitHub and OpenAI’s current dominance to their ability to deploy massive AI models with billions of parameters via the cloud. Although using these large models incurs higher operational costs, Copilot has managed to offset expenses to some extent through techniques like request batching.

However, this approach has its drawbacks: a report from the Wall Street Journal indicated that in early 2023, Microsoft was losing over $20 per month for each GitHub Copilot user.

In comparison, TabbyML aims to simplify the deployment process with models trained on 1-3 billion parameters, even if this trade-off may lead to lower initial quality. “As computing costs decrease and open-source models enhance, the competitive advantage of GitHub and OpenAI will gradually decline,” Zhang asserted.