Open Coding Agents (SERA): Specializing Code Agents on Your Own Repository

Coding agents can now debug, refactor, and open pull requests — but most share the same limitation: they’re closed, expensive to train, and can’t be adapted to a private codebase. A closed model has never seen your internal APIs, data pipelines, or org conventions, and teaching it those has historically meant generating synthetic training data that was both hard and costly to produce.

Ai2 Open Coding Agents targets that gap with an open training recipe for building a coding agent on any codebase, personal or internal. The first model family in the release is SERA (Soft-verified Efficient Repository Agents), built on Qwen3, ranging from 8B to 32B parameters, and compatible with Claude Code out of the box.

What makes it cheap

Two techniques do the heavy lifting.

Soft-verified Generation (SVG). Training a coding agent normally means generating pairs of buggy and corrected code, then testing each pair to confirm the fix is actually correct — expensive infrastructure and expensive generation. SVG’s finding is that patches don’t have to be fully correct to be useful for training. Generating only partially-correct patches removes the need for exhaustive correctness testing, and the reported result is that this “soft-verified” data scales like fully-verified data.

A bug-type menu. Rather than hunting for real bugs, SERA draws from a taxonomy of 51 common bug patterns. Each function in a repository can seed several distinct bug prompts, so a repo with thousands of functions yields tens of thousands of varied training trajectories cheaply.

A third insight ties them together: good synthetic data should mirror how a developer works a problem, not the exact details of the correct fix. That makes it practical to generate training data for essentially any repository, and to specialize with a straightforward supervised fine-tuning (SFT) job instead of a bespoke RL pipeline and test harness.

Reported performance

On SWE-Bench Verified, SERA-32B is reported to solve 54.2% of problems, trained in roughly 40 GPU days on a small cluster of NVIDIA Hopper or RTX PRO 6000 Blackwell Server Edition GPUs. Under aligned inference conditions:

• At 32K context, SERA-32B scores 49.5% ± 1.9% — comparable to Devstral Small 2 (50.0% ± 1.3%) and GLM-4.5-Air (50.5% ± 1.3%).
• At 64K context, it reaches 54.2% ± 1.4%, within ~4.9 points of Devstral Small 2, despite being pure SFT trained no further than 32K tokens.

On teacher choice, GLM-4.6 yields the best numbers, but GLM-4.5-Air gets close at lower cost — the gap widens mainly in higher-compute regimes, so a cheaper teacher is often the better call for early iterations.

The main result: specialization beats size

The reason the method matters is repository-specific specialization. Tested on Django, SymPy, and Sphinx — the three largest repositories in SWE-Bench, together over 60% of its problems — specialized models trained on 8,000 synthetic trajectories per repo matched or beat their 100B+ teachers. At 32K context the specialized models scored 52.23% on Django and 51.11% on SymPy, against GLM-4.5-Air’s 51.20% and 48.89%.

The takeaway is that a 32B model fine-tuned to a specific codebase can match or surpass a general-purpose teacher more than three times its size — with lower memory, faster inference, and reduced operating cost. In one reported case, SERA-32B surpassed its 110B teacher (GLM-4.5-Air) on Django and SymPy after training on 8,000 samples at a cost of about $1,300.

Cost and hardware

The reproduction costs reported are the headline argument: roughly $400 of commodity cloud compute to match the previous best open-source result, and about $12,000 to rival top open-weight industry models of the same size. Against comparable methods, the release reports matching SWE-smith at 57× lower cost and SkyRL at 26× lower cost.

For inference, Ai2 worked with NVIDIA on accelerated deployment. Reported throughput: ~1,950 peak output tokens/sec in BF16 on 4×H100 at a 16K context window, ~3,700 tokens/sec at FP8, and ~8,600 tokens/sec on Blackwell 4×B200 in NVFP4.

What ships

The release is fully open — models, Claude Code integration, all generated training data, and the recipe to generate your own — with an inference server that launches in two lines of code. The pipeline is deliberately simple: standard SFT on trajectories, no custom RL infrastructure. That combination is the point: it brings agent specialization within reach of small teams and independent developers, not just well-funded labs.