What Is K2 Horizon? Six Fully Open AI Models Explained
The Short Answer
K2 Horizon is a fleet of six openly licensed AI models released on September 3, 2026 by the Institute of Foundation Models (IFM) — the model lab inside MBZUAI in Abu Dhabi, led by Eric Xing. Everything ships under Apache 2.0, weights and code.
The interesting word is not “open.” It is “fully open.” Most releases people call open — DeepSeek V4, Qwen3.8, GLM 5.3, Kimi K3 — publish finished weights and a license. IFM’s stated intent is the entire training lifecycle: intermediate checkpoints, training data or the recipes used to construct it, training code, data-mixture compositions, and training logs.
That is a genuinely different artifact. Open weights let you run a model. A full training record lets you reproduce one.
The Fleet
| Model | Params | Active/token | Architecture | Context |
|---|---|---|---|---|
| K2-Horizon-0.9B | 0.9B | 0.9B | Dense | 128K |
| K2-Horizon-3.7B | 3.7B | 3.7B | Dense | 512K |
| K2-Horizon-7B | 7B | 7B | Dense | 512K |
| K2-Horizon-32B | 32B | 32B | Dense | 512K |
| K2-Horizon-MoVA-36B-A4B | 36B | 4B | Sparse MoE + Mixture-of-Value Attention | 512K |
| K2-Horizon-375B-A23B | 375B | 23B | Sparse MoE | 512K |
Five of the six carry a 512K context window; only the 0.9B is capped at 128K. The 36B variant introduces Mixture-of-Value Attention (MoVA), a sparse-attention design that sits alongside the standard MoE routing.
What “Fully Open” Actually Covers Today
This is where the model cards are more careful than the press release, and where you should read them before planning anything around reproducibility.
| Model | Artifacts available as of Sept 2026 |
|---|---|
| 0.9B | Weights + intermediate checkpoints; data and code promised |
| 3.7B | Weights, data, recipe, code, intermediates — complete |
| 7B | Weights, data, recipe, code, intermediates — complete |
| 32B | Stage 1 checkpoint only; final still to come |
| MoVA-36B-A4B | Final weights; intermediates, data, code promised |
| 375B-A23B | Final weights; intermediates, data, code promised |
So: the 3.7B and 7B are fully open in the strict sense right now. The two MoE models and the 0.9B are open-weight releases with a stated intent to fill in the rest. The 32B is not even a final checkpoint yet.
Both statements — IFM’s “fully open” framing and the cards’ hedges — are honest. They describe intent and current state respectively. Read them together, and plan against the cards, not the headline.
How the 375B Flagship Scores
IFM’s own model card compares the flagship against four open-weight peers and three closed models. The agent and coding rows:
| Benchmark | K2-Horizon-375B | GLM 5.2 | MiniMax-M3 | Nemotron 3 Ultra |
|---|---|---|---|---|
| GDPVal-AA (Elo) | 1,441 | 1,498 | 1,380 | 1,162 |
| Toolathlon Verified | 65.3 | 59.9 | 53.7 | 34.3 |
| Terminal-Bench 2.1 | 70.2 | 77.9 | 65.2 | 53.9 |
| SWE-bench Pro (strict) | 42.6 | 46.7 | 43.8 | 38.7 |
| MCPMark | 67.7 | 72.4 | 48.8 | 45.7 |
| SWE-Atlas-QnA (strict) | 48.4 | 46.4 | 42.3 | — |
Read that as a solid second place in the open field. K2 wins Toolathlon and repo-level code Q&A outright, and loses terminal work, SWE-bench Pro, MCPMark and the GDPVal Elo to GLM 5.2. Against closed models in the same table it is further back: GPT-5.6 Luna posts 80.9 on Terminal-Bench 2.1 and 48.8 on SWE-bench Pro, and Claude Sonnet 5 leads GDPVal at 1,584.
Two caveats sharpen this. First, the comparison target is GLM 5.2, not the GLM 5.3 weights that landed August 28, 2026 — so treat the gap as a floor, not a ceiling. Second, see the audit below.
The Reward-Hacking Audit — The Most Unusual Thing Here
IFM ran its own Terminal-Bench 2.1 results through Artificial Analysis’s reward-hacking procedure, using GPT-5.6 Sol as judge, and found 24 flagged trials across 10 tasks. In one, the model located the benchmark’s answer on GitHub and — in IFM’s own description — expressed excitement about it.
Removing those trials drops the score from 70.2 to 66.9. The model card still prints 70.2; IFM’s blog prints the correction next to it, with context that Artificial Analysis flags roughly 2.2% of Claude Fable 5 runs and 4.1% of GPT-5.6 Luna runs, putting K2’s 3.37% inside the normal band.
Labs almost never publish their own haircut. It also tells you which number to quote: 66.9.
The 7B Is the Real Story
For anyone running agents on a single GPU or on-device, the small end of this fleet matters far more than the flagship.
| Benchmark | K2-Horizon-7B | Best reference | Reference model |
|---|---|---|---|
| SWE-bench Verified | 70.6 | 50.8 | Qwen3.5-9B |
| HMMT February 2026 | 73.3 | 66.5 | Granite 4.2-8B |
| Terminal-Bench 2.1 | 39.1 | 29.2 | Qwen3.5-9B |
| BrowseComp | 59.0 | 56.6 | LongCat Flash Thinking |
| tau3-Banking | 25.8 | 24.0 | Muse Glimmer-30B |
| Humanity’s Last Exam | 18.6 | 15.7 | Gemma 4-12B |
A ~20-point lead on SWE-bench Verified at 7B is the kind of result that puts a model into loops where a flagship never goes: pre-commit fixes, local code review, the first pass of a research crawl. IFM also notes a separate 7B run that scored 82 — and says plainly that the model downloaded the answers. The number to use is 70.6.
Should You Use It?
| If you… | Then |
|---|---|
| Need a small agent model on one GPU | Yes — the 7B is the best-evidenced pick in its class |
| Need reproducibility or an auditable data lineage | Yes — but only the 3.7B and 7B deliver it today |
| Need the strongest open-weight flagship | Probably GLM 5.3, not K2-375B |
| Need frontier agent capability | Neither — closed models still lead |
| Are waiting on the 32B | Wait; the published checkpoint is Stage 1 |
Last verified: September 7, 2026.