The Institute of Foundation Models at MBZUAI in Abu Dhabi has released the K2-Horizon family — six fully open models ranging from 0.9B to a 375B-parameter mixture-of-experts model with 23B active parameters — under Apache 2.0. The models were first released on 3 September 2026, and the technical paper was submitted to arXiv on 16 November 2026 (arXiv:2611.21003).
The technical points of note: a new attention design called MoVA, softplus gating, a decoding method called Uno reported at about 2.2x speed, and a 512K-token context window. The flagship model scores 47 on the Artificial Analysis Intelligence Index.
The most unusual detail in the release is not a capability figure. The institute self-audited one of its own coding-benchmark figures and published the corrected, lower number — 70.2% down to 66.9%. Publicly correcting your own result downward is rare in model releases, and it is worth recording as part of the story.
Key facts
| Item | Details |
|---|---|
| Models | K2-Horizon family — six fully open models |
| Range | 0.9B to 375B-parameter mixture-of-experts (23B active parameters) |
| Licence | Apache 2.0 |
| Released by | Institute of Foundation Models, MBZUAI (Abu Dhabi) |
| First release | 3 September 2026 |
| Technical paper | Submitted to arXiv 16 November 2026 (arXiv:2611.21003) |
| Attention design | MoVA (new) |
| Gating | Softplus gating |
| Decoding method | Uno, reported at about 2.2x speed |
| Context window | 512K tokens |
| Flagship score | 47 on the Artificial Analysis Intelligence Index |
| Self-audit | One coding-benchmark figure corrected downward by the institute itself: 70.2% → 66.9% |
What this means, and what to watch
A fully open family spanning 0.9B to 375B parameters, under Apache 2.0, gives researchers and builders a wide range of model sizes to work with under one architecture family — including a 512K-token context window, which is large by current standards. The technical choices (MoVA attention, softplus gating, the Uno decoding method) will be judged by how they hold up in independent testing. And the self-audit sets a standard worth watching: if more labs correct their own figures downward in public, benchmark numbers across the field become easier to trust.
A note on verification
The facts above are verified from the sources named in the research file: the MBZUAI press release, the arXiv paper (arXiv:2611.21003), chainaiwire, NYU Shanghai RITS and emergent.sh. The figures are as reported; no figure has been added or adjusted.
FAQs
What is the K2-Horizon family?
Six fully open models from the Institute of Foundation Models at MBZUAI (Abu Dhabi), ranging from 0.9B to a 375B-parameter mixture-of-experts model with 23B active parameters, released under Apache 2.0.
What are the headline technical features?
A new attention design called MoVA, softplus gating, a decoding method called Uno reported at about 2.2x speed, and a 512K-token context window. The flagship scores 47 on the Artificial Analysis Intelligence Index.
When were the models released?
The models were first released on 3 September 2026. The technical paper was submitted to arXiv on 16 November 2026 (arXiv:2611.21003).
What is the self-audit story?
The institute checked one of its own coding-benchmark figures, found it overstated, and published the corrected, lower number — 70.2% down to 66.9%. A public correction of your own results is rare in model releases.
Source: MBZUAI press release; K2-Horizon technical paper (arXiv:2611.21003, 16 November 2026), as reported by chainaiwire, NYU Shanghai RITS and emergent.sh. Last checked: 21 November 2026.
Comments (0)
Share your question or experience about this syllabus. Your comment will appear after review.
No comments yet — be the first to share your thoughts!