The Institute of Foundation Models (IFM), backed by UAE research university MBZUAI, released K2 Horizon on 3 September 2026 — a fleet of six fully open AI models ranging from 0.9B to 375B parameters, all under the Apache 2.0 license permitting unrestricted commercial use.
K2 Horizon is not a single model but a coordinated fleet. All six sizes share core architecture, vocabulary, training methodology, and evaluation tooling, enabling teams to develop on a smaller model and deploy on a larger one with minimal code changes. IFM designed the fleet specifically for agentic tasks — tool use, multi-step reasoning, and long-horizon workflows — making it a direct option for enterprises building AI agents for business processes.
Six Models, One Shared Architecture
K2 Horizon ships six model sizes tuned for different deployment environments, according to BigDATAwire’s coverage of the release:
- 0.9B — designed for wearables such as watches and glasses
- 3.7B and 7B — designed for mobile phones; both set new state-of-the-art results at their respective scales on reasoning, tool use, and agentic benchmarks
- 32B (dense) — designed for local hosting and on-premise servers
- 36B-A4B (sparse MoE) — efficient on-premise option using a mixture-of-experts architecture
- 375B — the largest model in the fleet; released as the world’s largest fully open AI model, meaning weights, code, and training data are all publicly available
The 0.9B model also achieves state-of-the-art results at its scale on reasoning, tool use, and agentic tasks, according to IFM’s evaluation data cited by BigDATAwire.
What “Fully Open” Means Here
K2 Horizon releases weights, code, training data, and methodologies — all four components under Apache 2.0. Most open-weight releases from Meta (Llama 4) and Alibaba (Qwen 3.8) provide weights and code but withhold training data. IFM’s decision to include training data supports reproducibility and independent auditing, which matters for regulated industries evaluating AI governance risk.
The Apache 2.0 license permits unrestricted commercial use with no royalties or data-sharing requirements. Enterprises that self-host K2 Horizon models keep their data on-premise and avoid cloud provider dependencies — a compliance advantage for financial services, healthcare, and government procurement contexts.
Enterprise Use Cases and the Fleet Design Advantage
The shared vocabulary and architecture across all six sizes create a routing capability that single-model releases cannot match. A business can run a 7B model for low-latency screening, escalate complex tasks to a 32B or 36B-A4B model on an on-premise server, and run the most demanding workflows on 375B — all within the same inference stack. For enterprises building AI agents for business workflows, the ability to swap model sizes without rewriting integration code reduces deployment risk significantly.
K2 Horizon competes with open-weight AI models from Meta, Alibaba, and Mistral. Its differentiator is the fleet design: six coordinated sizes with shared tooling, plus full training data disclosure — a combination no other lab has shipped as of September 2026.
Our Take
K2 Horizon is the most deployable open-weight release of 2026 for enterprises that want on-premise control without vendor lock-in. The fleet design — six sizes, one vocabulary, engineered to be mixed in the same system — solves a real operational problem: most organizations need different model sizes for different workload tiers, and routing between incompatible models is expensive to maintain. IFM has built the routing in from the start. Business buyers evaluating self-hosted AI agents should put K2 Horizon on their shortlist alongside the leading AI tools for business in 2026.
For Context: Open AI Models and Agentic Workflows
- AI agents for business workflows — how multi-agent coordination works at the enterprise level
- 12 Best AI Agents for Business Tasks — where K2 Horizon-powered agents fit in the current landscape
