Flide accelerates decision inference through exact state forking.
Flide is an open-source inference engine for decision models, using exact state forking to cut repeated computation and claim up to 9.5x speed.
Key Takeaways
- Flide is an open-source inference engine for Jev-like decision models that reduces duplicated computation by serving related questions from shared state.
- Its core technique, exact state forking, lets multiple decision branches start from a common state while preserving independent evaluation semantics.
- The authors report up to 9.5× faster inference than vLLM and SGLang in their tests, but present this as a best-case result.
- Flide currently supports 30+ open-source decision models on Decision Index 0.4.1, with more models planned.
- The project is installable via `pip install flide` and offers public code, documentation, and a paper for community feedback.
Flide authors have released Flide, an open-source inference engine aimed at developers and researchers working with Jev-like decision models. The project offers a way to serve multiple related questions from shared state, aiming to reduce duplicated computation while keeping decision results consistent.
What Flide reports about performance
The authors say Flide reaches up to 9.5× faster inference than vLLM and SGLang in their testing. The result is a best-case claim, not a promise that every model, workload, or deployment will see the same gain. They also report support for 30+ open-source decision models on Decision Index 0.4.1, with more models planned.
Why exact state forking matters
Decision models often evaluate several related questions from a similar starting point. If each question recomputes the same state from scratch, serving time and resource use can rise quickly. Exact state forking is the project’s core idea: branches can start from a common state while preserving the decision semantics of independent evaluation.
How to access the project
The package can be installed with one command:
`pip install flide`
The code is available at https://github.com/OmniJev/Flide, the documentation at https://docs.flide.org/, and the paper at https://arxiv.org/abs/2610.42148. The authors invite feedback, questions, and ideas from the community.
What to watch next
The key follow-ups are whether independent users can reproduce the 9.5× speed claim across real workloads, how the 30+ model support expands beyond Decision Index 0.4.1, and whether documentation and community feedback support stable use.
Meta description: Flide authors released an open-source inference engine for Jev-like decision models, claiming up to 9.5× faster inference than vLLM and SGLang in tests.
Tags: Flide, inference, decision models, open-source, vLLM
Featured image: abstract friendly illustration of branching decision paths converging into a single fast pathway, no people, no logos, no text.

Comments (0)