Introducing OpenReasoningEngine, an open-source test-time-compute engine that can be used with any OpenAI-compatible model. Image input, function calling, basic continual learning, + more. This is an early experiment — there are issues that will need to be ironed out. Thread:

The engine guides the model to think step-by-step, and at each step, allows it to use code interpreters, web search, etc. to iterate solutions, test approaches, and gather info before responding. So, when it finally responds, the answer is more likely to be accurate.

One of the initial goals is to continually learn from experience, so the system can improve over time. Users can save successful reasoning chains to the memory bank, and the AI can use these memories to plan approaches to solving new problems.

A number of composable reasoning strategies have been implemented, (basic thought search, MoA), with more to come. Try each to see how well they work for your specific use-case. If the community comes up with compelling new approaches, they can be added to improve the system.

And soon as we get open-source TTC models (like DeepSeek R1), ideally we can connect them to this system, allowing them to use tools as they think to improve performance.

Think of this as a basic toolkit to allow the community to explore test-time-compute. It won’t be better at everything, but for many use-cases, it can help. Don’t expect magic.

I’m excited to share this with the community — it's a great base for building custom reasoning engines, or for use on its own. Consider this v0.1 - just a starting point. Each part of the system can be massively improved — I (and hopefully others) will iterate it over time.

A huge thank you to the folks who gave great advice, feedback, and helped build out and test the initial system. @stevenic @alecvxyz @josh_bickett @aidangomez @mlejva

If you have ideas for making it better or want to contribute, I'd love to hear from you! Let's build something truly useful together. GitHub:

I hope you will be vindicated, if socialmedia-heads weren't dumb NPCs they would understand that you single-handedly frontrunned the whole step-by-step reasoning AI trend a là o1 that is still widely going strong, you just slopped it a bit but the point still stands

Thank you, Matt, for your transparency AND this experimental engine. Can't wait to give it a spin, along with our developers.

Thanks for the kind words Shaun. Excited for you to try it!