The Institute of Foundation Models (IFM) at Abu Dhabi’s MBZUAI released K2 Horizon on September 3: a fleet of six fully open models spanning 0.9B to 375B parameters, under Apache 2.0, with intermediate checkpoints, training data or recipes, code, configs and fine-grained training logs attached to every one (ifm.ai, Reuters).
🔍 THE BOTTOM LINE
This is the most complete open model release to date — not just weights, but the entire training lifecycle. If “open weights” was this year’s argument, K2 Horizon makes it look half-finished: you can now study not only what a frontier-scale model knows, but how it was taught.
What “fully open” actually means here
Most open-weight releases hand you the final checkpoint and call it a day. IFM has opened everything: pretraining data (or detailed construction recipes where licensing blocks redistribution), intermediate checkpoints across training, the LLM360 fully-open principle it has followed since 2023, mixture compositions, training code, and loss logs fine-grained enough to show where instabilities appeared.
The fleet itself is six models sharing one architecture and training recipe: 375B-A23B (MoE, enterprise), a dense 32B and a sparse 36B-A4B for local workstations, and 7B, 3.7B and 0.9B for on-device use. The 0.9B, 3.7B and 7B models claim new state-of-the-art results at their size classes, with the 0.9B scoring above 48 on AIME 2026 — remarkable for a model small enough to run quantised on a watch (ifm.ai).
The technical bets worth noticing
Two things in the release stand out beyond the leaderboard claims.
First, MoVA (Mixture-of-Value Attention) extends expert sparsity into the attention layer itself, not just the feed-forward layers. The 36B-A4B model activates only ~4B parameters per token yet lands nearly at the dense 32B’s performance. If that ratio holds up under independent evaluation, it changes the economics of local inference.
Second, Uno Diffusion — a LoRA adapter that learns to generate token blocks in parallel while keeping the autoregressive model’s output distribution frozen. IFM claims it’s lossless, meaning same answers, roughly three times faster. Speculative decoding usually needs a separately trained draft model; this is a plug-in adapter. Reuters reports the models are live on Hugging Face and major inference providers now.
There’s also a quietly interesting training detail: roughly 17% of the ~20-trillion-token pretraining corpus was explicit reasoning trajectories, with about 10 trillion synthetic tokens overall. Reasoning built in from pretraining, not bolted on in post-training.
The geopolitical layer
The UAE is now a serious player in open AI infrastructure, and this lands in the same quarter as Washington easing export controls on UAE AI chips. A Gulf state building sovereign, reproducible frontier models — and giving them away under Apache 2.0 — is a deliberate counter-position to both closed US labs and Chinese open-weight players like Moonshot and Alibaba. Free models are not free to produce; someone is underwriting this, and sovereign AI strategy is the visible reason.
What it means for New Zealand
This is close to the ideal release for a small ecosystem. The 7B and 3.7B models are designed for local hardware — the kind of machines NZ developers, students and small studios actually own. And because the full training recipe is open, a university course can now walk students through the entire process, not just inference. The 32B class is realistic for a single high-memory workstation.
❓ FAQ
Is K2 Horizon really better than closed frontier models? At the top end, no — the 375B-A23B ranks “among the top models” in its class but isn’t claiming to beat GPT-5.6 or Claude Opus 5. The genuinely newsworthy claims are the small-model SOTA results at 0.9B–7B, where open models now lead outright.
Can I run K2 Horizon on my own machine? Yes. The 7B and 3.7B targets phones and on-device use, the 32B suits a high-RAM workstation, and all six ship with quantisation support. Weights are on Hugging Face under Apache 2.0, so commercial use is permitted.
How is this different from DeepSeek or Qwen releases? Scope. Those releases open the weights and some training details; K2 Horizon opens checkpoints, data recipes, code, configs and per-step training logs across the whole lifecycle — the first release to do so through agentic post-training.
Why does a UAE university give away frontier-scale models? Sovereign AI strategy. Fully open releases build international influence and developer mindshare the way closed APIs can’t — the same logic that has driven China’s open-weight push, now pursued from the Gulf.
🔍 THE BOTTOM LINE
The open-vs-closed debate has mostly been about final weights. K2 Horizon shifts the argument to something harder to dismiss: full reproducibility, from data mixture to loss curves, at sizes from smartwatch to data centre. Whether the benchmark claims survive independent scrutiny matters less than the precedent — the definition of “open” in AI just moved, and it moved because a state-backed institute decided to spend the compute to make it move.
📰 Sources
- Institute of Foundation Models — K2 Horizon announcement
- Reuters — Abu Dhabi AI institute releases fully open-source models
- PR Newswire — IFM launch release
