Thank you :) Swift Models hit 2.2 million+ downloads / Early Access to New Models, Free Compute for Researchers

Hey everyone,

Jovan from UkisAI (Swift Qwen) here!

For those who don't know us, UkisAI is a small lab making tiny frontier LLMs, tools and datasets (+doing it open-source!). I'm one of the guys running it aka I train the models and post on Reddit.

Our first open-source release is Swift, a series of reasoning-efficient LLMs. It is proof of how penalizing pathological overthinking patterns inside of various LLMs can bring their token usage down -58.3% and speed x1.95 without losing accuracy if RL-ed correctly afterwards by not training them to think shorter directly but rather to think more efficiently. You can find Swift 27B here as well as Swift Flash Next here we have GSQ-RCO quants (kudos to ISTA-DAS) and uncensored variants (thank you community).

It's honestly unbelievable to me that our models have crossed 2M downloads... The team and I had a deal that we shall do a toast (drinks) after we hit 100k, and I'm honestly not sure how to celebrate now but in the meantime I want to thank everyone who contributed to our models, be it the independent benchmarks, quantizations, finetunes or just using them. Without all of you guys, we would have had no way to continue our work, and now with the downloads rolling in we are more than happy (and paid hah) to continue training new models and as of recent making other tools for local AI users. On that matter, I'm sharing two things with you today:

1. We are making a Discord community so we can talk to Swift users more easily, get your thoughts and ideas on things as well as test new models and tools we've been working on :)

The first 100 people to join will get early access to our:

- Unreleased Swift models (we have trained Swift GLM 5.3 Flash and Swift 9B and are looking for early testers before putting it on HuggingFace!)

- UkisAI Code (Codex modified and optimised for local models, we use it internally to have remote-control with open-source models, better browser use, /loop etc)

- Swift.cpp (inference engine, we do all of our training and coding internally via local models so we made an engine that's optimized for Swift models specifically and runs up to 30% faster on our hardware)

After this the community shall stay open for everyone but we are still figuring out the mechanics of early-access so that part shall be invite-only for the time being. This shouldn't matter to most people as all of the stuff testers get access to will be open-source regardless if it's any good.

To join Early Access (max 100 people): https://discord.gg/S2nNfQNFxh

Regular link to join: https://discord.gg/XvX9J8nbkJ

1. We're also making the UkisAI Research Support Program

- We want to provide free compute, LLM APIs and early access to our datasets for amazing people experimenting with building models of their own or working on new things with Swift models.

As it's our first time making this we can't estimate our capacity right away so there is not a specific number of individuals we can help with research but if this sounds interesting to you please message me on Discord and I'll see to it.

End note:

We are big believers in local AI and that open-source will win replacing all the proprietary cloud models for personal use, but even as users ourselves we don't have all the ideas and solutions to make that happen. This is why we need the community to help us know what to build.

Please share your model requests, tools you need, problems you have with local AI regardless of if you've been using Swift models or need more of them - they are just one of the things we need to make to let local AI be better than the cloud.

Let's cook!