Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More
Qodo, an AI-driven code quality platform formerly known as Codium, has announced the release of Qodo-Embed-1-1.5B, a new open source code embedding model that delivers state-of-the-art performance while being significantly smaller and more efficient than competing solutions.
Designed to enhance code search, retrieval, and understanding, the 1.5-billion parameter model achieves top-tier results on industry benchmarks, outperforming larger models from OpenAI and Salesforce.
For enterprise development teams managing vast and complex codebases, Qodo’s innovation represents a leap forward in AI-driven software engineering workflows. By enabling more accurate and efficient code retrieval, Qodo-Embed-1-1.5B addresses a critical challenge in AI-assisted development: context awareness in large-scale software systems.
Why code embedding models matter for enterprise AI
AI-powered coding solutions have traditionally focused on code generation, with large language models (LLMs) gaining attention for their ability to write new code.
However, as Itamar Friedman, CEO and co-founder of Qodo, explained in a video call interview earlier this week: “Enterprise software can have tens of millions, if not hundreds of millions, of lines of code. Code generation alone isn’t enough—you need to ensure the code is high quality, works correctly, and integrates with the rest of the system.”
Code embedding models play a crucial role in AI-assisted development by allowing systems to search and retrieve relevant code snippets efficiently. This is particularly important for large organizations where software projects span millions of lines of code across multiple teams, repositories, and programming languages.
“Context is king for anything right now related to building software with models,” Friedman said. “Specifically, for fetching the right context from a really large codebase, you have to go through some search mechanism.”
Qodo-Embed-1-1.5B provides performance and efficiency
Qodo-Embed-1-1.5B stands out for its balance of efficiency and accuracy. While many state-of-the-art models rely on billions of parameters—OpenAI’s text-embedding-3-large has 7 billion, for instance—Qodo’s model achieves superior results with just 1.5 billion parameters.
On the Code Information Retrieval Benchmark (CoIR), an industry-standard test for code retrieval across multiple languages and tasks, Qodo-Embed-1-1.5B scored 70.06, outperforming Salesforce’s SFR-Embedding-2_R (67.41) and OpenAI’s text-embedding-3-large (65.17).
This level of performance is critical for enterprises seeking cost-effective AI solutions. With the ability to run on low-cost GPUs, the model makes advanced code retrieval accessible to a wider range of development teams, reducing infrastructure costs while improving software quality and productivity.
Addressing the complexity, nuance and specificity of different code snippets
One of the biggest challenges in AI-powered software development is that similar-looking code can have vastly different functions. Friedman illustrates this with a simple but impactful example:
“One of the biggest challenges in embedding code is that two nearly identical functions—like ‘withdraw’ and ‘deposit’—may differ only by a plus or minus sign. They need to be close in vector space but also clearly distinct.”
A key issue in embedding models is ensuring that functionally distinct code is not incorrectly grouped together, which could cause major software errors. “You need an embedding model that understands code well enough to fetch the right context without bringing in similar but incorrect functions, which could cause serious issues.”
To solve this, Qodo developed a unique training approach, combining high-quality synthetic data with real-world code samples. The model was trained to recognize nuanced differences in functionally similar code, ensuring that when a developer searches for relevant code, the system retrieves the right results—not just similar-looking ones.
Friedman notes that this training process was refined in collaboration with NVIDIA and AWS, both of whom are writing technical blogs about Qodo’s methodology. “We collected a unique dataset that simulates the delicate properties of software development and fine-tuned a model to recognize those nuances. That’s why our model outperforms generic embedding models for code.”
Multi-programming language support and plans for future expansion
The Qodo-Embed-1-1.5B model has been optimized for the top 10 most commonly used programming languages, including Python, JavaScript, and Java, with additional support for a long tail of other languages and frameworks.
Future iterations of the model will expand on this foundation, offering deeper integration with enterprise development tools and additional language support.
“Many embedding models struggle to differentiate between programming languages, sometimes mixing up snippets from different languages,” Friedman said. “We’ve specifically trained our model to prevent that, focusing on the top 10 languages used in enterprise development.”
Enterprise deployment options and avail
Qodo is making its new model widely accessible through multiple channels.
The 1.5B parameter version is available on Hugging Face under the OpenRAIL++-M license, allowing developers to integrate it into their workflows freely. Enterprises needing additional capabilities can access larger versions under commercial licensing.
For companies seeking a fully managed solution, Qodo offers an enterprise-grade platform that automates embedding updates as codebases evolve. This addresses a key challenge in AI-driven development: ensuring that search and retrieval models remain accurate as code changes over time.
Friedman sees this as a natural step in Qodo’s mission. “We’re releasing Qodo Embed One as the first step. Our goal is to continually improve across three dimensions—accuracy, support for more languages, and better handling of specific frameworks and libraries.”
Beyond Hugging Face, the model will also be available through NVIDIA’s NIM platform and AWS SageMaker JumpStart, making it even easier for enterprises to deploy and integrate into their existing development environments.
The future of AI in enterprise software dev
AI-powered coding tools are rapidly evolving, but the focus is shifting beyond code generation toward code understanding, retrieval, and quality assurance. As enterprises move to integrate AI deeper into their software engineering processes, tools like Qodo-Embed-1-1.5B will play a crucial role in making AI systems more reliable, efficient, and cost-effective.
“If you’re a developer in a Fortune 15,000 company, you don’t just use Copilot or Cursor. You have workflows and internal initiatives that require deep understanding of large codebases. That’s where a high-quality code embedding model becomes essential.”
Qodo’s latest model is a step toward a future where AI isn’t just assisting developers with writing code—it’s helping them understand, manage, and optimize it across complex, large-scale software ecosystems.
For enterprise teams looking to leverage AI for more intelligent code search, retrieval, and quality control, Qodo’s new embedding model offers a compelling, high-performance alternative to larger, more resource-intensive solutions.
