IFM K2 Horizon: Six Fully Open AI Models for Enterprise
On September 3, 2026, the Institute of Foundation Models (IFM) released K2 Horizon: a fleet of six AI models spanning 0.9 billion to 375 billion parameters, all under Apache 2.0, all shipped with training code, data, intermediate checkpoints, and logs. It is the largest fully open-source model launch in AI history, and it changes the calculus for enterprise teams evaluating self-hosted and sovereign AI deployments.
This is not another “open-weight” release where you get the weights and nothing else. K2 Horizon is reproducible from the ground up, which matters for regulated industries, auditors, and any organization that needs to verify what went into the model it is running.
What IFM Released and Why It Matters
IFM, a global AI research lab launched in May 2025 by MBZUAI (Mohamed bin Zayed University of Artificial Intelligence) with labs in Abu Dhabi, Silicon Valley, and Paris, released six models at once. Each one ships with:
- Final model weights under Apache 2.0
- Pre-training and post-training code
- Training data (directly, or reproducible construction recipes where redistribution is restricted)
- Intermediate training checkpoints across the full training lifecycle
- Fine-grained training logs and evaluation results
“Rather than release a single model, we’re releasing an entire fleet at once: six models, from one small enough to run on a watch to a flagship built for enterprise reasoning,” said Hector Liu, director of IFM’s Silicon Valley lab. “Every model is built to compete with the best open models at its size, and every one ships with the weights, code, training data and methodology behind it.”
For enterprise AI teams, this level of transparency has two immediate implications. First, it removes the “black box” objection that regularly stalls procurement and legal reviews. Second, it creates a foundation for fine-tuning and specialization with full knowledge of what the base model learned and how.
The Six Models: From Edge to Enterprise
The K2 Horizon fleet covers every deployment profile from wearables to data-center-scale reasoning. Each model is designed to compete with the best open-weight models at its size class.
| Model | Architecture | Active Params | Primary Use Case |
|---|---|---|---|
| K2 Horizon 0.9B | Dense | 0.9B | Watches, glasses, edge devices |
| K2 Horizon 3.7B | Dense | 3.7B | Phones, on-device fine-tuning |
| K2 Horizon 7B | Dense | 7B | Local assistants, coding, general-purpose |
| K2 Horizon 32B | Dense | 32B | Workstations, on-premise servers |
| K2 Horizon 36B-A4B | Sparse MoVA | ~4B | Efficient server inference |
| K2 Horizon 375B-A23B | Sparse MoE | ~23B | Enterprise reasoning, long-horizon agents |
The 36B-A4B model introduces a new architecture IFM calls Mixture-of-Value-Attention (MoVA). It activates only about 4 billion parameters per token while drawing on 36 billion stored parameters, giving teams a cost-effective inference profile with capacity well above its compute cost.
The three small models (0.9B, 3.7B, 7B) set new state-of-the-art results at their respective scales on reasoning, math, coding, and agentic benchmarks. The 0.9B model scores above 48 on AIME 2026, a result that would have required a model many times larger just one generation ago.
Benchmark Results: Where K2 Horizon Stands
IFM benchmarked K2 Horizon 375B-A23B against both open-weight and closed frontier models. The flagship is competitive with models considerably larger by total parameter count and compares favorably to mid-tier closed models on agentic tasks.
| Benchmark | K2 Horizon 375B | Inkling (xhigh) | MiniMax-M3 | GPT-5.6 Luna (max) | Claude Sonnet 5 (max) |
|---|---|---|---|---|---|
| Toolathlon (agentic tool use) | 65.3 | 45.5 | 53.7 | 67.5 | 71.6 |
| Terminal-Bench 2.1 (agentic terminal) | 70.2 | 55.1 | 65.2 | 80.9 | 80.5 |
| SWE Bench Pro (software engineering) | 42.6 | 43.1 | 43.8 | 48.8 | n/a |
| MCPMark (MCP tool use) | 67.7 | 51.2 | 48.8 | 66.9 | 65.3 |
| GPQA Diamond (graduate science QA) | 87.3 | 87.2 | 92.9 | 91.1 | 91.1 |
| AA-Omniscience Non-Hallucination | 74.7 | 32.0 | 82.0 | 7.0 | 61.0 |
One result worth flagging: the non-hallucination rate on AA-Omniscience. GPT-5.6 Luna scores 7.0 and Claude Sonnet 5 scores 61.0, while K2 Horizon 375B scores 74.7. This benchmark, which measures the rate at which a model correctly declines questions outside its knowledge rather than confabulating, is directly relevant for enterprise deployments where incorrect confident answers create downstream risk.
IFM also conducted a reward-hacking audit on Terminal-Bench results using Artificial Analysis’s methodology. The audit flagged 3.37% of passing trials (for comparison, Claude Fable 5 flags at 2.2% and GPT-5.6 Luna at 4.1%), resulting in a corrected accuracy of 66.9% (down from 70.2%). This level of methodological transparency is unusual in AI model launches and is precisely what enterprise due-diligence processes require.
Enterprise Deployment: What You Actually Need
Deploying K2 Horizon requires understanding the distinction between “open” and “cheap to run.” The flagship 375B-A23B, while legally and technically open, requires substantial GPU infrastructure. IFM’s validated serving configuration uses an eight-H200 node with tensor and expert parallelism, BF16 precision, and FlashAttention-3.
For teams without that infrastructure, the practical enterprise options are:
API through inference partners. Compass, Cerebras, and Nebius offer K2 Horizon 375B-A23B APIs today, which lets enterprise teams access flagship performance without running their own cluster.
On-premise with the 32B or 36B-A4B models. The 32B dense model runs on a high-end workstation or a modest on-premise server. The 36B-A4B MoVA model offers similar or better performance with lower compute cost per token.
Edge and mobile with 7B and below. The 7B and 3.7B models ship with quantization support for phones and edge devices. The 0.9B is small enough for watches and glasses.
All sizes are available through Hugging Face, with day-zero support from vLLM, SGLang, and Ollama. Hardware compatibility extends to NVIDIA, AMD, and Cerebras accelerators.
What This Means for Enterprise AI Strategy
Three strategic implications for organizations building or buying AI systems:
Full openness changes the vendor relationship. When you can reproduce training, inspect every data decision, and verify benchmark claims against your own evaluation harness, you have real leverage in procurement and in customization. This is categorically different from a vendor that releases weights and publishes a technical report.
Sovereign and regulated deployments become more defensible. Organizations in financial services, healthcare, defense, and government that need to justify AI decisions to regulators can point to verifiable training artifacts. This matters more than raw benchmark scores for many compliance frameworks.
The fleet model simplifies architecture decisions. Because all six K2 Horizon models share the same architecture family, training methodology, and deployment tooling, teams can prototype on the 3.7B model and promote to 32B or 375B without changing their inference stack. This is a meaningful operational advantage over assembling a model portfolio from multiple vendors.
For enterprise AI teams evaluating open-weight models as part of a broader AI-native transformation strategy, K2 Horizon raises the bar on what “open” should mean. Earlier open releases like Tencent Hunyuan HY3 offered Apache-licensed weights but not the training pipeline. K2 Horizon closes that gap, positioning IFM as a serious alternative to both closed frontier providers and the more limited open-weight releases that have dominated the market since 2025.
Enterprise teams comparing self-hosted options against closed frontier models such as Claude Opus 5 will find K2 Horizon 375B competitive on agentic tasks, with a considerably better transparency profile and no commercial license restrictions.
What to Watch Next
IFM describes the launch as the beginning of a staged rollout. The final model weights are available now, but some announced artifacts, including intermediate checkpoints and training code for certain models, are still being staged for release. Teams evaluating K2 Horizon for production should verify which artifacts are actually available before committing to a deployment plan.
Independent benchmark validation is also underway across several AI evaluation groups. IFM’s self-reported numbers are consistent with the methodological rigor of the release, but enterprise procurement decisions should wait for third-party corroboration on the specific benchmarks most relevant to the target workload.
IFM has signaled continued development. The K2 roadmap shows K2 V2 (December 2025), K2 Think (January 2026), and K2 Horizon (September 2026) as milestones, suggesting regular release cadence with increasing capability and openness. For enterprise teams that need both frontier performance and full auditability, K2 Horizon sets a new reference point.