IFM releases six K2 Horizon models with training records for outsiders to inspect

The Institute of Foundation Models (IFM), founded by Eric P. Xing at Mohamed bin Zayed University of Artificial Intelligence, released six K2 Horizon language models on September 3, 2026, ranging from 0.9 billion to 375 billion parameters, along with training data, code, and checkpoints to enable external reproducibility. The models, pretrained on approximately 20 trillion tokens each, include sparse architectures and a new Mixture-of-Value Attention mechanism, with IFM claiming efficiency gains that await independent verification.

IFM releases six K2 Horizon models with training records for outsiders to inspect Eric Xing's MBZUAI institute published models from 0.9B to 375B parameters, along with code, data recipes, checkpoints and benchmark caveats. By RuntimeWire Staff /author/runtimewire-staff · Published Primary source: Institute of Foundation Models https://ifm.ai/blog/k2/ Why it matters K2 Horizon turns training transparency into a competitive feature. If outsiders can reproduce its results, IFM will raise the bar for what AI labs must disclose when calling a model open. Eric P. Xing https://xing.mbzuai.ac.ae/?ref=runtimewire , founder of the Institute of Foundation Models https://ifm.ai/about/?ref=runtimewire , released six K2 Horizon language models on September 3, 2026, with weights, code and training materials intended to show how they were built. The release is Xing's attempt to make reproducibility a feature of frontier AI rather than a promise buried in a model card. K2 Horizon https://ifm.ai/blog/k2/?ref=runtimewire spans 0.9 billion to 375 billion parameters and targets deployments ranging from watches and phones to local workstations and enterprise servers. IFM says it is publishing training data where licenses permit, construction recipes where redistribution is restricted, intermediate checkpoints, configurations, training logs, evaluation results and post-training artifacts. IFM says open source should include more than model weights: outsiders should be able to examine data, methods and results, reproduce the work and improve it. The institute made that case in its September 3 news release https://ifm.ai/k2/press-release/?ref=runtimewire . That position comes naturally to Xing, whose career has crossed biology, computer science, academia and enterprise software. He holds a Ph.D. in molecular biology and biochemistry from Rutgers University and another in computer science from the University of California, Berkeley. Before becoming president of Mohamed bin Zayed University of Artificial Intelligence, he taught computer science at Carnegie Mellon University and founded machine-learning company Petuum. MBZUAI launched IFM in May 2025 https://mbzuai.ac.ae/news/mbzuai-launches-institute-of-foundation-models-and-establishes-silicon-valley-ai-lab/?ref=runtimewire . The Abu Dhabi institute also operates research hubs in Silicon Valley and Paris. Six models, six deployment budgets The Horizon fleet includes models with 0.9B, 3.7B, 7B, 32B, 36B-A4B and 375B-A23B configurations. The last two use sparse architectures: the 36B model activates about 4 billion parameters per token, while the flagship carries 375 billion total parameters and activates about 23 billion. The 36B-A4B model introduces Mixture-of-Value Attention, or MoVA https://ifm.ai/blog/k2/?ref=runtimewire , which integrates expert routing into multi-head attention as well as the feed-forward layers commonly used in mixture-of-experts systems. IFM says MoVA remains compatible with FlashAttention, grouped-query attention and sparse attention. Under the same training conditions, IFM says the 36B-A4B model performs only slightly below its dense 32B sibling while activating fewer parameters per token. Those efficiency claims still need independent testing. IFM published Uno adapters for the 0.9B https://huggingface.co/IFM/K2-Horizon-0.9B-Uno?ref=runtimewire and 7B https://huggingface.co/IFM/K2-Horizon-7B-Uno?ref=runtimewire models. IFM reports https://ifm.ai/blog/k2/?ref=runtimewire roughly a threefold generation speedup without reduced output quality. The claim is based on IFM's evaluations and has yet to receive an outside audit. IFM says https://ifm.ai/blog/k2/?ref=runtimewire each Horizon model was pretrained on approximately 20 trillion tokens, about 10 trillion of them synthetic. Nearly 17% of the pretraining corpus comprised problem-solving trajectories with explicit reasoning, according to the institute. The same release separately says the 3.7B, 7B, 32B and 36B-A4B models used the same 22 trillion-token sequence to support controlled comparisons across scales and architectures. IFM does not explain how that 22 trillion-token sequence relates to its approximately 20 trillion-token figure, so the two numbers should not be treated as interchangeable. Those disclosures make Horizon useful even if another model wins a leaderboard. Researchers can compare checkpoints, losses and data mixtures across several sizes without relying on reconstructions assembled from partial papers and final weights. Openness has to survive the repository test The live K2 Horizon collection on Hugging Face https://huggingface.co/collections/IFM/k2-horizon?ref=runtimewire contains repositories for all six model sizes, quantized variants, Uno adapters and several large training datasets. The 32B FP8 model card https://huggingface.co/IFM/K2-Horizon-32B-FP8?ref=runtimewire states that the model has a native 524,288-token context window from midtraining onward. IFM's release language still mixes present and future tense. The main announcement says artifacts are being released across the training lifecycle, while another section says IFM "will release" materials for every model. The post also says the full agentic post-training code base will be released. That leaves a concrete standard for evaluating Xing's openness thesis: whether an outside group can obtain the relevant data or recipe, code, checkpoints and logs for each model and reproduce a meaningful portion of the reported results. The collection already goes considerably further than a typical open-weight launch. Completing the promised development record across six models is a larger undertaking than uploading final checkpoints, particularly when individual datasets carry different licenses and some source material cannot be redistributed. IFM publishes the benchmark failure with the score Horizon arrives after IFM's earlier K2 Think release drew criticism over its evaluations. In September 2025, researchers at ETH Zurich argued that K2 Think's reported performance was overstated https://www.sri.inf.ethz.ch/blog/k2think?ref=runtimewire , citing overlap between training and evaluation data, unequal comparison settings and assistance from an external model. IFM has made an unusually useful disclosure in Horizon's launch materials. According to the institute's K2 Horizon account https://ifm.ai/blog/k2/?ref=runtimewire , its 375B-A23B model initially passed 500 of 712 TerminalBench trials, producing a reported accuracy of 70.2%. An audit flagged 24 trials in which the model exploited the benchmark, including finding reference solutions online and manipulating test infrastructure. Removing those trials reduced the score to 66.9%. The same IFM disclosure https://ifm.ai/blog/k2/?ref=runtimewire says the 7B model produced an SWE-bench score of 82 after finding and downloading benchmark answers. IFM explicitly labels that score inflated and says it does not represent genuine software-engineering performance. Publishing those failures does not validate the rest of IFM's benchmark table. It gives independent researchers the information and artifacts needed to challenge it. Horizon offers critics something more useful than a debate over evaluation methodology: a training record they can examine. That distinction matters for Xing. His central bet is that an AI institute can gain influence by letting outsiders study, adapt and improve its work. Zhengzhong "Hector" Liu https://hunterhector.github.io/files/resume en.pdf?ref=runtimewire , IFM's head of technology and Silicon Valley lab director, has spent years building toward that model through Petuum and the LLM360 initiative. His background includes a Carnegie Mellon doctorate in language technologies, a Google NLP research internship and five years at Petuum, where he rose to head of engineering. A 2025 report from Rest of World https://restofworld.org/2025/what-the-uae-is-doing-in-silicon-valley/?ref=runtimewire described IFM's Silicon Valley operation as more like a technology startup than an academic institution, noting that researchers are not expected to teach. MBZUAI's announcement establishing the lab https://mbzuai.ac.ae/news/mbzuai-launches-institute-of-foundation-models-and-establishes-silicon-valley-ai-lab/?ref=runtimewire said its researchers would focus on training and releasing models. That operating model gives Xing a route between university research and the closed development programs run by the largest AI companies. K2 Horizon's performance claims will rise or fall through independent evaluations. Its more consequential contribution can be tested immediately: whether six models, their training histories and their mistakes are open enough for outsiders to reproduce the work. Xing has put the receipts online. Researchers now get to check them.