CONFIRMED

The Institute of Foundation Models (IFM), a research institute affiliated with MBZUAI university, released K2 Horizon on September 3, 2026, a family of six fully open artificial intelligence models, with weights, source code, training data and complete methodology available under an Apache 2.0 license. The release was confirmed by HPCwire, PR Newswire, Moor Insights & Strategy and CryptoBriefing, which described the package as the largest fully open source AI model release ever made.

Six models spanning watches to local servers

K2 Horizon consists of six models ranging from 0.9 billion to 375 billion parameters: 375B-A23B and 36B-A4B, both sparse, the dense 32B model, and the smaller 7B, 3.7B and 0.9B versions. According to IFM, the 0.9 billion parameter model was designed for highly constrained environments such as smartwatches and glasses, while the 3.7B and 7B models bring advanced capability to phones and other on-device applications. The dense 32B model and the sparse 36B-A4B model target local servers and on-premise infrastructure, for teams that want to run AI without depending on a third party cloud.

Fully open: weights, code and training data

What sets K2 Horizon apart from other releases labeled open source is that it does not just publish model weights, but also the code, the data used in training and the complete methodology, all under an Apache 2.0 license. That lets researchers and developers inspect, reproduce and adapt the models for their own projects, something most frontier AI labs still do not do, even when they call their models open.

Top tier performance in every size class

According to IFM, K2 Horizon delivers top tier performance on reasoning, mathematics, coding and agentic tasks in every size class, with the 0.9B, 3.7B and 7B models setting a new state of the art at their respective scales. Those numbers come from the institute itself, not an independent benchmark, but the full availability of weights and code lets any developer validate the claims by running their own tests, something closed benchmarks do not allow.

What it means for the open AI market

K2 Horizon arrives at a moment when American companies are increasingly willing to adopt open models, many of them coming from institutes and labs outside the United States, as a cheaper alternative to closed models from OpenAI and Anthropic. A family of six models spanning wearable devices to local servers, all under a permissive license, widens the options for anyone who wants to run AI without depending on a single cloud or a single vendor.

Why it matters for Brazilian agencies and SMBs

For Brazilian customer service and marketing automation agencies, fully open models like K2 Horizon expand the option of running AI locally, on their own server or on domestic cloud, reducing dependence on foreign vendors and, in some cases, inference cost at scale. The smaller models, from 0.9B to 7B, are especially relevant for anyone who wants to embed AI in devices or apps without paying for an external API on every call. It is worth testing the 32B model or the sparse 36B-A4B in an internal pilot before migrating sensitive workloads, and watching whether the developer community independently confirms the performance numbers IFM published.

Sources

HPCwire, Institute of Foundation Models Releases Fully Open K2 Horizon Models with Weights, Code and Training Data: https://www.hpcwire.com/aiwire/2026/09/03/institute-of-foundation-models-releases-fully-open-k2-horizon-models-with-weights-code-and-training-data/ | Institute of Foundation Models, K2 Horizon Press Release: https://ifm.ai/k2/press-release/ | Moor Insights & Strategy, MBZUAI's IFM Launches 6 K2 Horizon Frontier Models, Doubles Down on Openness: https://moorinsightsstrategy.com/mbzuai-ifm-launches-6-k2-horizon-frontier-models-doubles-down-on-openness-analyst-insight/