Continuum is a unified runtime for LLM and ML programs. It executes token generation and tensor computation through a shared intermediate representation (IR), so caching, dispatch, and interoperability are handled by one system instead of ad-hoc glue.

Quick Start

Install:

python -m pip install continuum-ai

Use:

import continuum

Run a reproducible benchmark validation:

PYTHONPATH=python python scripts/benchmarks/run_examples.py | python scripts/benchmarks/validate_outputs.py

Try the main example:

examples/01_research_agent.py

Why Continuum

One runtime model: token ops and tensor ops run in one executable graph.
Backend-agnostic cache reuse: reusable state handles enable cross-call prefix reuse without backend-specific app code.
Capability-driven dispatch: backends are selected by declared capabilities (tensor, token, cache).
Explicit tensor interoperability: cross-backend conversions are explicit and type-tagged.
Native-first core: C++ engine and ABI-focused design for long-term extensibility.

What Is Implemented

C++ execution engine with IR interpreter
KV cache index with canonical prefix normalization
Azure backend for real network execution
libtorch backend for tensor/training execution
MLX backend for Apple-native tensor paths

Current Status

v1 release hardening in progress
CIR schema lock with serialization conformance (schema/cir.fbs)
Linux and macOS CI matrix with coverage gates and fuzz workflow
PyPI packaging under continuum-ai (import path remains continuum)

Documentation

Python API docs: https://ct.rithul.dev/python/
C++ API docs: https://ct.rithul.dev/cpp/

Build docs locally:

# Python docs
python -m venv .venv-docs
. .venv-docs/bin/activate
pip install sphinx furo breathe
PYTHONPATH=python sphinx-build -b html docs/api/python docs/api/python/_build
 
# C++ docs
doxygen Doxyfile

Local outputs:

docs/api/python/_build/index.html
docs/api/cpp/html/index.html

Community

Contributing guide: CONTRIBUTING.md
Code of Conduct: CODE_OF_CONDUCT.md
Security policy: SECURITY.md
Support guide: SUPPORT.md
Governance: GOVERNANCE.md

Quick contributor setup:

pip install pre-commit
pre-commit install
pre-commit run --all-files
pytest

Citation

If Continuum helps your work, cite it as:

@software{continuum2026,
  title        = {Continuum: Unified Runtime for Token and Tensor Programs},
  author       = {Kamesh, Rithul and Contributors},
  year         = {2026},
  url          = {https://github.com/rithulkamesh/continuum},
  version      = {1.0.0}
}