VectorArc/avp-python

Implements vocabulary-mediated projection to transfer computation across incompatible model families without retraining, automatically negotiating between full KV-cache transfer (same model), ~6 KB projected hidden states (cross-model), or text fallback. Integrates with HuggingFace, Ollama, llama.cpp, and vLLM engines, plus LangChain, CrewAI, and AutoGen frameworks—enabling 2-3x faster multi-agent pipelines while maintaining or improving accuracy on reasoning benchmarks.

Available on PyPI.