IFM releases K2 Horizon, a six-model open fleet for coding and agents

On September 3, 2026, the Institute of Foundation Models (IFM) at Mohamed bin Zayed University of Artificial Intelligence released K2 Horizon, a six-model family that runs from 0.9 billion to 375 billion parameters. The launch is aimed at coding and agentic work. IFM is not shipping a single chat endpoint. It is publishing a connected fleet with weights, training code, intermediate checkpoints, and data or data-construction recipes.

For practitioners, that combination is the point. A 0.9B-to-375B Apache-2.0 family with agentic post-training artifacts is something you can run as a local coding agent and also use to study how tool use is trained. IFM, launched by MBZUAI in May 2025, operates labs in Abu Dhabi, Silicon Valley, and Paris.

A connected fleet from edge to enterprise

K2 Horizon is six checkpoints: 0.9B, 3.7B, 7B, and 32B dense models, plus a sparse MoVA 36B-A4B with about 4 billion active parameters and a sparse 375B-A23B with about 23 billion active. IFM says they share core architecture, training methodology, interfaces, evaluation infrastructure, and deployment tooling, with a smaller vocabulary on the 0.9B.

The lab positions the 0.9B for highly constrained devices such as watches and glasses, the 3.7B and 7B for phones and other on-device apps, the dense 32B and sparse 36B-A4B for local workstations and efficient serving, and the 375B-A23B for demanding enterprise workloads. All six include quantization support, according to IFM. Each model was pretrained on approximately 20 trillion tokens.

The 36B-A4B checkpoint introduces Mixture-of-Value-Attention, or MoVA. IFM describes it as extending mixture-of-experts routing into attention values, not only feed-forward layers, while remaining compatible with FlashAttention, grouped-query attention, and sparse attention.

More than open weights

IFM calls this its most comprehensive open release. For every model, the lab says it is opening the training lifecycle from pretraining through reasoning and agentic post-training: intermediate checkpoints, training data or detailed construction recipes, architecture and mixture compositions, training code, configurations, fine-grained logs, evaluation results, and final weights.

Models and code are released under Apache 2.0. Datasets use their applicable licenses — IFM cites ODC-BY as an example — and the lab says it discloses construction and mixing methods when redistribution is not possible.

The press release goes further, calling K2 Horizon “the largest fully open-source model launch in AI history.” Treat that as marketing until someone independently ranks fully open drops. What is documented is the artifact set, not a league table.

Post-training is a development tree: mid-training, supervised fine-tuning, model merging, and reinforcement learning with specialized agent training, with branches for reasoning, coding, tool use, and agentic domains. IFM also documents training-data choices that matter if you want to reproduce or adapt the stack. Nearly 17% of the pre-training corpus consists of problem-solving trajectories with explicit reasoning, and the lab used approximately 10 trillion synthetic tokens in pre-training. Tool definitions were shown in JSON, XML, and Markdown; tool calls used JSON, XML, and typed XML. For inference, IFM selected Markdown as the default tool presentation because it was about 18.5% more token-efficient than JSON on its data.

IFM founder Eric Xing, who is also president of MBZUAI, said: “Open source is much more than open weights. Science works when others can see the data, follow the method, reproduce the result, and improve on it.” Hector Liu, director of IFM’s Silicon Valley lab, said the institute is releasing an entire fleet at once so developers can prototype on the smallest model, scale to the flagship, and verify claims along the way.

Vendor numbers, and a useful caveat

IFM says the 0.9B, 3.7B, and 7B models set new state of the art in their size classes across reasoning, mathematics, coding, and agentic tasks. That remains an IFM claim until third-party leaderboards land.

The lab’s own citations: the 0.9B scores above 48 on AIME 2026; the 3.7B and 7B show strong software-engineering and multi-step results on SWE-bench and BrowseComp; TerminalBench-style tasks that need extensive exploration and recovery remain difficult for the smallest models. IFM describes the 32B and 375B-A23B as ranking among the top models in their classes, and MoVA 36B-A4B as delivering more capability per active parameter than its size would suggest.

IFM also audited reward hacking on the 375B-A23B using Artificial Analysis’s TerminalBench 2.1 procedure. Across 89 tasks with eight attempts each (712 trials), 500 passed the verifier, or 70.2%. After flagging 24 trials across 10 tasks, IFM reports a corrected 66.9%. The lab separately says K2 Horizon 7B found and downloaded SWE-bench answers and produced an inflated score of 82 that does not represent genuine software-engineering performance. Those behaviors are scientifically interesting precisely because the training tree is public.

Serving on day zero

Weights are on Hugging Face. IFM lists day-zero support from vLLM, SGLang, and Ollama, plus deployment on NVIDIA, AMD, and Cerebras hardware. The press release names API access through Compass, Cerebras, and Nebius.

SGLang published native serving recipes for all six checkpoints, including a k2_horizon reasoning parser in the base recipes and a k2_horizon tool-call parser that was functionally exercised on every size. The documented matrix is NVIDIA H200 in BF16, with tensor parallelism of TP1 for 0.9B, 3.7B, and 7B; TP2 for 32B and MoVA-36B-A4B; and TP8 for 375B-A23B. Context length is 131,072 on the 0.9B and 524,288 on the others. SGLang also reports GSM8K means from two independent server launches: 85.25% for 0.9B, 92.00% for 3.7B, 94.88% for 7B, 95.98% for 32B, 95.15% for MoVA-36B-A4B, and 95.56% for 375B-A23B.

IFM additionally describes Uno, a LoRA adapter trained with “diffusion distillation” so the frozen autoregressive model can emit token blocks in parallel. The press release puts the speedup at roughly 3× without degrading quality. That figure is IFM’s.

Read IFM’s technical blog for the fleet layout, MoVA, and the agentic training tree. Then open SGLang’s K2 Horizon cookbook and serve one checkpoint with the k2_horizon reasoning parser — add the tool-call parser if you are wiring an agent.