K2 Horizon's phone-sized models are fully open — and you can't run them yet
MBZUAI's IFM released a 3.7B and a 7B model built for phones on September 3, but upstream llama.cpp does not support the architecture yet, so nothing built on stock llama.cpp — Pocket AI included — can load them today.
On September 3, 2026, the Institute of Foundation Models (IFM) — the research arm of Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), in Abu Dhabi — released K2 Horizon: six models from 0.9B to 375B parameters, every one of them Apache 2.0, with weights, training code, data recipes, intermediate checkpoints and evaluation logs all published alongside. Two of the six sizes — 3.7B and 7B — are explicitly built for phones. That is exactly the kind of release this blog exists to flag.
Here is the catch: as of this writing, you cannot load them in llama.cpp, and therefore not in Pocket AI, and not in most of the other local-model apps on your phone either. The GGUF files exist. The runtime to read them does not, yet.
What actually shipped
K2 Horizon is a family, not one model. The dense line is 0.9B, 3.7B, 7B and 32B parameters. Alongside it sit two sparse mixture-of-experts models: a 36B model with roughly 4B active parameters per token, and a 375B flagship with roughly 23B active. The whole fleet was trained on roughly 20 trillion tokens.
"Fully open" is IFM's own term for it, and it means more than the usual "open weights" release: alongside the model files they published the training code, the data construction recipes, intermediate checkpoints from training, and evaluation logs — the kind of transparency that traces back to the 2023 LLM360 project rather than to a typical lab model-card drop. Restricted datasets get documented recipes instead of raw redistribution, where licenses do not allow the data itself to be shared.
One naming note worth flagging before you search for this: "K2" is also the name of Moonshot AI's unrelated Kimi K2, a much larger model with a very different release philosophy. They share nothing but a name.
The two sizes that matter here
K2-Horizon-3.7B and K2-Horizon-7B are both dense, decoder-only models with a native 524,288-token (512K) context window from the midtraining stage onward. The smallest model in the family, K2-Horizon-0.9B, is aimed lower still — IFM lists it at 1.08B stored parameters once you count embeddings, with a 131,072-token (128K) context using YaRN RoPE scaling. That gap between the marketing number (0.9B) and the actual stored size (1.08B) is the same kind of rounding this blog has flagged before — check the real number, not the name on the box.
IFM's own claim is that the 7B is the strongest model under 10B parameters, and the 3.7B the strongest under 4B, on reasoning, math, coding and agentic benchmarks. We have not reproduced any of those numbers ourselves, and no independent third-party benchmark run turned up in our search — so treat that as the lab's claim, not a verified result, until someone outside IFM publishes numbers.
Why it doesn't run yet
GGUF files for all four dense sizes are already up on Hugging Face, uploaded by IFM themselves (IFM/K2-Horizon-3.7B-GGUF, IFM/K2-Horizon-7B-GGUF, and so on), and a community GGUF conversion of the 36B MoE model exists too. A GGUF file existing is not the same as a GGUF file loading, though — llama.cpp needs code that understands a model's specific architecture to read its tensors, and that code has to be written and merged before any app built on llama.cpp, Pocket AI included, can open the file.
For K2 Horizon, that code is not merged. The tracking discussion on the llama.cpp repository (opened the same day as the release, September 3) describes a draft implementation from an MBZUAI-IFM fork proposed upstream, not yet part of mainline llama.cpp. It also flags a more specific snag: the 3.7B and 7B models were built against a different Transformers version than the one llama.cpp currently pins, which is its own source of compatibility breakage independent of the architecture-support question.
Practically, that means: no llama.cpp-based app can load K2-Horizon-7B-Q4_K_M.gguf today without building against IFM's own unofficial fork instead of mainline llama.cpp — something no shipping consumer app does, Pocket AI included, because pinning to a fork means losing every upstream fix and update until the fork is abandoned or merged.
Should you wait for it
If you care about full transparency — training data recipes, intermediate checkpoints, code you can actually audit rather than a model card with a benchmark table — K2 Horizon's 3.7B and 7B are worth bookmarking. That level of openness is rare even among labs that call their own releases "open." The context window (512K, natively, not stretched with a rope hack after the fact) is also genuinely useful at that parameter count.
But there is nothing to download onto your phone from this release today. Not "the file is too big" or "the quality isn't there" — the runtime literally cannot read it yet. Skip it until llama.cpp support lands, then it is worth a second look. We will follow up here when it does.
Sources
- MBZUAI Institute of Foundation Models, K2 Horizon launch announcement, September 3, 2026
- llama.cpp discussion #28308, "K2 Horizon: Pre-release llama.cpp support," opened September 3, 2026