SSM Native

SSM‑Native Frontier AI for Software‑Driven Systems

Modern software development needs models that understand structure, semantics, and developer workflows. SSM Native is a research initiative at TU Darmstadt building models that generate high-quality software more efficiently.

We develop long-context, sample efficient model with superior length extrapolation capability (at least 32x). We aim for performance parity with transformers while achieving substantial efficiency gains: 10-12x during training, ~20x during inference, and 12-15x higher throughput. We will ensure performance and scalability with sufficient efficiency for local deployments. Our initial focus is secure and correct code generation in software engineering and later extend to wider application across various domains such as Robotics, Digital Circuit Design.

GitHub Explore Research

Interpretability Driven Model Design

Efficient Model

Workflow Integration

Benchmarking

News

Latest updates from the SSM Native project

Loading news updates...

See all news

Research

Four interconnected pillars of SSM-native framework

Interpretability Driven Model Design

We improve upon our SSM model architecture based on the behaviour of the learned representation of the model. We build SSM architectures that capture code semantics, application structure, developer intent, and repository level context.

Interpretable Models
State Space Models for code
Structure-Aware Generation

Efficient Model

We develop models which are highly efficient than the current LLMs. Our model consumes less than 20% of the GPU memory than what current models consume, which makes them ideal for local deployment. Our model provides at least 12-15x faster generation.

Latency Reduction
Less memory consumption
Developer Efficiency

Workflow Integration

Our models connect directly to development workflows via IDE tools, code review assistants, and API generation pipelines. This makes SSM Native useful not just for research, but for real software teams aiming to speed up implementation and reduce cognitive load.

IDE Assistance
Code Review Support
API Generation

Evaluation & Benchmarking

We measure model performance on software production metrics such as correctness, maintainability, and task completion speed. Our benchmarking work validates that SSM Native improves end-to-end software generation, not just token-level accuracy.