With the release of Ai2’s open coding agents, developers have a new method for writing and testing software that promises to slash costs.
Coding agents allow engineering teams to automate debugging, refactoring, and even the submission of pull requests. However, most high-performance agents are proprietary, expensive to train, and unable to interact securely with private codebases.
Relying on closed models means sending intellectual property to external APIs, often without the model understanding the specific conventions, internal APIs, or data pipelines that define an organisation’s architecture. The Allen Institute for Artificial Intelligence (Ai2) has released a new method aimed at resolving this tension, lowering the barrier to entry for bespoke coding agents.
The primary obstacle to training custom coding agents has been the difficulty of generating high-quality synthetic training data from private repositories. This process is typically cost-prohibitive for smaller labs or independent teams. Ai2’s new release, the ‘Open Coding Agents’ family, introduces a training method that reduces these costs dramatically.
According to the institute, reproducing the performance of the previously best open-source model now costs approximately $400 in compute. For performance rivalling top-tier industry models, the cost rises to $12,000. This reduction in capital requirements suggests that small teams can now afford to fine-tune models on their own infrastructure, rather than relying on generic hyperscale providers.
Soft-verified generation explained
The core innovation driving this efficiency is a technique Ai2 calls Soft-verified Efficient Repository Agents (SERA). Standard approaches to synthetic data generation rely on creating pairs of incorrect and corrected code, which the agent uses to learn patching. Traditionally, these examples undergo rigorous testing to ensure absolute correctness, a process requiring complex infrastructure.
Ai2’s researchers found that training patches do not need to be fully correct to be instructive. Under the SERA methodology, the system generates synthetic training data using patches that are only partially correct. This approach, termed ‘Soft-verified generation’ (SVG), eliminates the necessity for heavy test harnesses while scaling effectively.
To prevent the model from learning only a narrow set of fixes, the system utilises a taxonomy of 51 common bug patterns. By applying these patterns to functions within a repository, the method generates tens of thousands of varied trajectories from a single codebase. The result is training data that mirrors a developer’s workflow rather than just the precise details of a correct code block.
Benchmarking the performance of Ai2’s open coding agents
In evaluations, the effectiveness of this streamlined approach appears evident. The 32-billion parameter model, SERA-32B, solves 54.2 percent of problems on the SWE-Bench Verified benchmark. This performance exceeds prior open-source models of comparable size and context length.
When tested against proprietary heavyweights, the results remain competitive. At a 32k token context window, SERA-32B achieves a resolve rate of 49.5 percent, placing it within a narrow margin of Devstral Small 2 (50%) and GLM-4.5-Air (50.5%). This is particularly notable given that SERA relies purely on supervised fine-tuning (SFT) without the complex reinforcement learning (RL) pipelines used by many competitors.
On NVIDIA Blackwell systems using NVFP4 precision, the model scales to approximately 8,600 output tokens per second. Even on existing H100 clusters, using BF16 precision yields around 1,950 tokens per second.
Advantage for private codebases
The most immediate application for developers is the ability to specialise models for internal software. General-purpose models often struggle with proprietary stacks because they lack exposure to the specific logic and conventions of the codebase.
Data from Ai2 suggests that a smaller open model trained on a specific repository can outperform a much larger generalist model. For instance, after training on just 8,000 samples, SERA-32B surpassed the 110-billion parameter GLM-4.5-Air teacher model on repositories like Django and SymPy.
This “teacher-student” inversion allows organisations to deploy efficient 32B models that behave with the competence of 100B+ parameter systems within their specific domain. This reduces memory requirements and operational costs while keeping data strictly within the organisation’s control.
The release includes the full training recipe, models, and data, designed to be launched with minimal configuration. It is also compatible out of the box with Claude Code, allowing for easier integration into existing developer toolchains.
For teams considering adoption, the pathway involves generating targeted synthetic data from their own repositories followed by a standard supervised fine-tuning job. This avoids the need to build elaborate reinforcement learning environments for every new task setting.
The democratisation of this capability means that agentic coding is no longer the exclusive domain of well-funded research labs. Whether for a small independent developer or a mid-sized enterprise, the ability to create a bespoke coding agent for a few hundred dollars changes the economics of software automation.
See also: Microsoft’s engineers are treating AI coding tools as standard practice
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events. Click here for more information.
Developer is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.