LLM360

A truly open reasoning model building on top of their previous line of models.

LLM360, a project from MBZUAI, is back with their fully open-source model series. This model is a 70B dense model, and they release the whole data, from pre-training (12T tokens) to SFT, which they generated using GPT-OSS 120B at all three reasoning levels. They also release multiple checkpoints from the various stages of training. We expect a lot more from them specifically and the growing fully-open model community in 2026!

MBZUA, mostly known for their open LLMs, released their first reasoning model built on top of Qwen2.5 32B, therefore making it not truly open. This was branded by the UAE as their first major foray into open models. Their technical report goes into more detail on the training, the data used, as well as some ablation studies - the key differentiation of the model and demo is the use of inference-time scaling methods potentially similar to GPT Pro or DeepThink.

There are around a handful of truly open models, which share the data and intermediate checkpoints during the training alongside the final artifacts. LLM360 has released fully open 7B models in the past and has now released a 65B open model, competitive with Llama 2.

A 65B "fully open-source" model. This model reaches similar performance to Llama 2 70B and uses less compute (only 1.4 trillion tokens).