IFM (MBZUAI) Open Models: Model List, Licenses and Articles

September 26, 2026November 11, 2026

About IFM (MBZUAI)

IFM (Institute of Foundation Models) is an AI research lab launched in May 2025 by MBZUAI (Mohamed bin Zayed University of Artificial Intelligence) in the United Arab Emirates. According to its IFM organization on Hugging Face, it operates from Abu Dhabi, Silicon Valley and Paris and is dedicated to “open and independent development of frontier-class foundation models".

IFM stands out for “fully open" models that release not only weights but training data, recipes and training code. As LLM360 it previously released models under a policy of publishing all checkpoints, code and data.

What It Releases

• K2-Horizon is its latest family, with sizes from 0.9B to 375B (September 2026): dense models at 0.9B, 3.7B, 7B and 32B, the 36B-A4B MoVA, and a 375B-A23B MoE. According to the 32B model card, the native context is 512K tokens, and the training data, recipe and training code will be made public.
• It releases intermediate checkpoints too, so that capability changes can be studied across training. Note that the results on the 32B model card are for stage 1 of training; stage 2 results will follow.
• LoopedLM is a set of research checkpoints for “looped" language models that pass through the same layers repeatedly (published September 28, 2026). They are the weights behind the experiments of the paper Towards Looped Models Done Right: Part I (comparing layer layout, input injection and recurrent-state organization, with MoE and dense variants at the 336B- and 500B-token points) and Part II (rethinking at fixed points). According to its model card, Part II’s huginn-s-parcae has about 303M parameters.
• K2-Type-0.9B is a “decision model" that returns probabilities for each option instead of generating text (September 26, 2026). Built on K2-Horizon-0.9B, it answers a state and typed questions in a single forward pass.
• Earlier families include the reasoning-focused K2-Think, the 65B K2, and the 7B Amber and Crystal. The organization card also lists Jais, an Arabic LLM, among IFM’s models.

Licenses

K2-Horizon and K2-Think are Apache-2.0, broadly usable including commercially. The guru 7B and 32B models, however, are non-commercial CC BY-NC 4.0, and MegaMath uses the license of its Llama 3.2 base. This summary is not legal advice; read each model’s license text before use.

Running Them Locally

• LoopedLM cannot be run with GGUF or Ollama. According to the model cards, the weights are in xLLM’s native format (BF16 Safetensors with an xLLM config.json), cannot be loaded with Hugging Face transformers, and need the code at ifm-ai/xllm-loop. They exist to reproduce the paper’s comparisons, not as everyday chat models. The license is Apache-2.0.

• K2-Horizon has official GGUF builds. IFM itself publishes GGUF files for 0.9B, 3.7B, 7B, 32B and MoVA-36B-A4B. The 32B and 375B have NVFP4 builds, and many sizes have FP8 builds.

• It is an easy family to pick a size from. On a consumer GPU the 7B or 32B GGUF, on a laptop the 0.9B or 3.7B, are the candidates. For which quantization fits which GPU, see the memory table on the family page.

Sources: the IFM organization card; the model card of IFM/K2-Horizon-32B; the model cards of IFM/LoopedLM-P2-huginn-s-parcae and IFM/K2-Type-0.9B; the model list of the IFM organization on Hugging Face (all as of 2026-10-03).

Our Coverage and Data

Local Model Watch has published 3 article(s) on IFM (MBZUAI)’s own models, 0 on third-party fine-tunes and quantized builds based on them, and 0 from its official blog. The model list below is collected by code from Hugging Face every day. Part of our publisher index; model families link to their family pages.

At a Glance

Recent Models on Hugging Face

The publisher lists 123 models in total on Hugging Face. Below are up to 15 of the most recent generative models (text, image, video, audio and so on), the scope of this site, per organization. Classifiers, feature extractors, research components and repositories without a model card are omitted, and the publisher’s own quantized builds and helper variants are folded into the row of the original model. Parameter counts are left blank for repositories that pack weights into integer types, where Hugging Face’s count is not the real parameter count.

From the Hugging Face API. Last changed 2026-11-11 (JST). Collected by code.

Licenses of Its Own Models (as recorded in our articles)

The license can differ from model to model even within one publisher. This is not legal advice; check each model’s license text before use.

By Model Family

Articles on Its Own Models

Last updated 2026-11-11 (JST). The explanation at the top of this page was written with the help of AI from the primary sources it cites. The tables and lists under “Our Coverage and Data" are assembled by code from our article log and the Hugging Face API.