SERA

SERA (Soft-Verified Efficient Repository Agents) is the first model in Ai2's Open Coding Agents initiative, an efficient method for training coding agents that enables rapid and affordable creation of agents specialized to private codebases. Using only supervised fine-tuning, SERA achieves 54% on SWE-bench Verified at 64K context (state-of-the-art for fully open-source models) while being 26x cheaper than reinforcement learning approaches. The key insight is that soft verification---comparing patches using line-level matching rather than executing test suites---produces training data of equal quality while removing the need for test infrastructure. This makes it practical to train a coding agent on your own codebase and deploy it with Claude Code. Learn more from the blog post or the SERA paper.

Why SERA?​

• Private codebase support: Specialize open-weight models to your proprietary code, encoding repository-specific patterns directly in the model weights
• No test infrastructure needed: Generate training data from any repository without requiring test suites or bug injection pipelines
• Cost-effective: Train a specialized coding agent for approximately $1,300
• Claude Code integration: Deploy your model and use it with Claude Code via sera-cli

Model variants​

The "-GA" variants are trained with GLM-4.5-Air, while SERA-32B uses GLM-4.6 as the teacher model for higher performance.

All models are available in the Open Coding Agents collection on HuggingFace.

Quick start​

The fastest way to try SERA is with Modal, which handles GPU provisioning and model deployment automatically.

# Install modal and sera-cli
uv tool install modal
uv tool install ai2-sera-cli

# Setup modal (prompts for account setup)
modal setup

# Deploy SERA to Modal and launch Claude Code
sera --modal

When you exit Claude Code, the Modal deployment automatically cleans up.

Using an existing endpoint​

If you have an existing vLLM endpoint:

# Install sera-cli
uv tool install ai2-sera-cli

# Set API key if required
export SERA_API_KEY=<your API key>

# Run with your endpoint
sera --endpoint <endpoint URL>

Team deployments​

For multi-user setups, create a persistent deployment:

# Deploy a persistent vLLM instance
deploy-sera --model Qwen/Qwen3-32B

# Share the endpoint URL and API key with team members
# Team members connect with:
SERA_API_KEY=<api-key> sera --endpoint <endpoint-url>

# Stop when done
deploy-sera --stop

Deploying private models​

If you have fine-tuned SERA on your own codebase, deploy the private model using a HuggingFace secret:

# Create a Modal secret with your HuggingFace token
modal secret create huggingface HF_TOKEN=hf_your_token_here

# Deploy your private model
deploy-sera --model your-org/private-sera-model --hf-secret huggingface

# Users connect with the provided endpoint and API key
SERA_API_KEY=<api-key> sera --endpoint <endpoint-url>

Training your own​

You can fine-tune SERA on your own codebase using the SERA training pipeline. The process generates synthetic coding agent trajectories from your repository using the Soft Verified Generation (SVG) method, then fine-tunes an open-weight model on those trajectories.

See the SERA training repository for full instructions.

Training data​

SERA is trained on over 200,000 synthetic coding agent trajectories generated using the Soft Verified Generation (SVG) method. See the SERA training data documentation for details on the data format and generation method.