SERA (Soft-verified Efficient Repository Agents)
Product · Developer tool by Ai2
Fast, accessible coding agents that adapt to any repo.
SERA is a family of open coding agents and a training framework that allows developers to build custom agents for private codebases at low cost. It offers state-of-the-art performance on SWE-Bench and high-throughput inference on NVIDIA infrastructure.
- Launched
- 2026-01-27
- Languages
- Python, TypeScript, Django
Category: Developer Tools
- Integrations
- Claude Code, NVIDIA accelerated infrastructure
- SWE-Bench Verified resolve rate (32B)
- 54.2%
- Peak throughput (4xH100 FP8)
- 3,700 tokens/sec
- Peak throughput (4xB200 NVFP4)
- 8,600 tokens/sec
- Context window
- 16k tokens
- Training cost (reproduction)
- ~$400 compute
- 54.2% SWE-Bench Verified resolve rate (SERA-32B)
- Up to 8,600 peak output tokens/sec on Blackwell 4xB200 systems
- 57x lower synthetic data costs compared to prior methods
- Out-of-the-box compatibility with Claude Code
Overview
SERA (Soft-verified Efficient Repository Agents) is a family of open coding agents and a training framework developed by Ai2 for software developers and organizations. It solves the problem of closed, expensive coding agents by providing an accessible method to build custom agents for private codebases, including internal APIs and specific organizational conventions. The SERA-32B model achieves state-of-the-art performance, solving 54.2% of SWE-Bench Verified problems while maintaining high inference speeds of up to 8,600 tokens per second on NVIDIA Blackwell systems. The framework is designed for efficiency, matching industry model performance at a fraction of the typical training cost through a Soft-verified generation (SVG) method. SERA is compatible with Claude Code out of the box and supports tasks ranging from code generation and debugging to maintenance and explanation.
Performance and Efficiency
- Cost-efficient training: The framework matches the performance of industry models at a fraction of the typical training and compute costs.
- Soft-verified generation: The training recipe uses Soft-verified generation (SVG) to reduce synthetic training data costs by up to 57x compared to prior methods.
Development and Integration
- Private codebase adaptation: Developers can specialize models to their own private data, including internal APIs and specific organizational conventions.
- Multi-task coding support: The framework supports tasks including code generation, code review, debugging, maintenance, and code explanation.
- Claude Code integration: SERA models are compatible with Claude Code out of the box for tasks like debugging and submitting pull requests.
- Single-line deployment: The system can be launched with a single line of code, making it accessible for those without LLM training experience.
Inference and Infrastructure
- H100 FP8 throughput: Running at FP8 precision on 4xH100 GPUs, SERA achieves approximately 3,700 peak output tokens per second.
- Blackwell NVFP4 throughputs: On Blackwell 4xB200 systems running in NVFP4, the model scales to around 8,600 peak output tokens per second.
- NVIDIA infrastructure optimization: The models support high-throughput inference in production environments through optimization for NVIDIA accelerated infrastructure.
- 16k context window: The system maintains a 16k context window for processing repository-level information.
Use cases
- Software engineers use SERA to debug and refactor code within their private organizational repositories using internal APIs.
- Development teams specialize SERA models to their specific engineering stack and conventions to improve code review quality.
- Researchers use the open training recipes and datasets to study agentic behavior and push the boundaries of open coding models.
- DevOps teams deploy SERA on NVIDIA Blackwell systems to achieve high-throughput inference for real-time code generation tasks.
Specifications
- Integrations
- Claude Code, NVIDIA accelerated infrastructure
- SWE-Bench Verified resolve rate (32B)
- 54.2%
- Peak throughput (4xH100 FP8)
- 3,700 tokens/sec
- Peak throughput (4xB200 NVFP4)
- 8,600 tokens/sec
- Context window
- 16k tokens
- Training cost (reproduction)
- ~$400 compute
- Training cost (high performance)
- Up to $12,000 compute
- Model variants
- 14B and 32B parameters
- Precision support
- BF16, FP8, NVFP4
How it works
- Access models and data: Developers can access the open weights for SERA-32B and SERA-14B, along with the open training datasets and recipes.
- Specialize to private codebases: The framework allows for specialization to private codebases using a simple, reproducible pipeline that doesn't require large-scale RL infrastructure.
- Deploy and integrate: The system can be launched with a single line of code and integrated with tools like Claude Code for active development tasks.
FAQ
What makes the SERA training framework efficient?
The training method uses Soft-verified generation (SVG) to reduce synthetic training data costs by up to 57x compared to prior methods like SWE-smith.
Can I train SERA on my own private codebase?
Yes, SERA is specifically designed to allow developers to specialize models to private data, including internal APIs and organizational conventions.
What kind of inference throughput can I expect?
SERA is optimized for NVIDIA accelerated infrastructure, reaching 3,700 tokens/sec on 4xH100 (FP8) and 8,600 tokens/sec on 4xB200 (NVFP4).
Added by AntSpace from public sources (allenai.org).