Reika
A coding agent CLI designed around small local models first
39 followers
A coding agent CLI designed around small local models first
39 followers
Reika is a coding agent CLI for local and hosted models, designed around local models first. It doesn't make a small model more intelligent but it provides a better experience using them day to day. Even if you're not part of the absolute range that Reika is designed for, using larger local models or hosted ones will still provide a solid experience out of the box.
Free
Launch Team
AWS Startup AdvisorPersonalized AWS guidance that meets you where you’re already building
Promoted
Reika was a project I never intended to make public because it was part of my own personal local AI stack. But as I worked on it to make it usable as a daily driver, I thought it would be nice to make it public for others to see and use.
I initially built it to see how much I could get the harness to make small models, especially at low quantization and context to not feel terrible to use. So while Reika doesn't solve the intelligence side, it tries to solve the overall experience when using small models at the absolute scale.
A lot of the testing and pain came through working on my M2 MacBook Air 16GB trying to run models like Qwen3.6 35B A3B and Qwen3.8 27B all day in agentic coding, maxing out the RAM and limits of my own machine. So the base of Reika comes from a legitimate source of truth.
Anyways, hope you all enjoy it :)
Hi Alex, can you explain more what you mean by "So while Reika doesn't solve the intelligence side, it tries to solve the overall experience"? What's an example of an experience you found improved by Reika?
Congrats on the launch today!
@sara_ford_goog
Hi Sara, thanks for the comment!
This example is more general but it's something I think is worth mentioning for an experience that Reika improves.
That would be providing the ability for someone with constrained hardware to be able to have a more smoother experience with agentic coding on the models they can run, on a limited context window and with low quantization. Because this combination often leads to disastrous model behavior such as doom looping within reasoning, tight context constraints, unnecessary tool call repetition, and task misalignment, the remainder of the core experience is up to the agent harness to normalize and smooth it out. It won't be perfect since there's a ceiling course, but getting most of the critically terrible parts to feel more natural in practice rather than an annoyance is what helps a lot in a daily driver scenario.
And it's also pretty much reducing the fear of missing out stress because your hardware can't run the models you want very well and instead leveraging what you have, given the right tools to make it happen. Therefore, Reika enables the other half to make your life a bit easier until you're ready to upgrade to better hardware and a local model setup. And in the high range, Reika scales nicely as you throw larger context windows and stronger models at it because it isn't heavily constrained by the guards presented on the absolute side that Reika is tuned against.
That's kind of the gist but hope that sort of makes sense.
@alex_dev6 Ah I see, okay, you're trying to mitigate the poor user experiences of doom looping, tool call repetition, etc. That makes sense. Thanks for all the details! Good luck with the launch!