New Free AI Model Clover 2 Beats Specialists at Their Own Game
This free AI could save you money and give better answers on tough tasks.
Clover 2 is a new AI model that works like a team of specialists. Instead of one big model trying to do everything, it has a base model and seven smaller 'expert' add-ons, each trained for a specific job: roleplay, general assistant, security, math, long-document handling, efficiency, and knowledge. A router decides which expert to use for each word (token) it generates, and can even fall back to the base model when no expert is needed. This is different from just merging all experts into one static model; it's a dynamic choice at every step.
The result is better performance on tests. For example, it scores higher on math problems (GSM8K, MATH-500) and long-document understanding (LongBench-v2). It also reduces the annoying writing tics that AI often has, like overusing certain phrases. So if you use AI for writing, math, or research, you get more accurate and natural results.
The team has made everything public: the base model, all seven experts, the router, and the code to run it. You can try it for free for a week at kitani.ai without a credit card. They also support zero data retention (ZDR) and have a clear privacy policy, so your data isn't stored.
One catch: if you want to run it on your own computer, it's not straightforward yet. There's no easy version for common local setups (no GGUF), and the recommended way is using vLLM, a more technical tool. So for now, it's best for those with some technical know-how or for using the hosted free trial.
- Clover 2 uses seven specialized experts and picks the best one for each word, improving accuracy on math, long documents, and creative writing.
- It's free to try for a week at kitani.ai with no credit card, and your data isn't stored (ZDR).
- Self-hosting is possible but requires technical setup; no easy local version yet.
Why It Matters
You get better AI answers for free, with less generic writing and more accurate math, without compromising privacy.