K2 Horizon: six open models that ship the training data

Most "open" model releases hand you weights and keep everything that made them. K2 Horizon does not. On September 3, the Institute of Foundation Models at MBZUAI published six models under Apache 2.0 — from 0.9B to 375B parameters — along with the code, the methodology, and the training data. For a small team, the size range matters more than the flagship: there is now a permissively licensed model for the phone, the laptop, and the server, trained the same way.

What actually happened

Per MBZUAI's announcement, the fleet is 0.9B for edge devices, 3.7B for on-device work, 7B for phones, 32B for local hosting, a 36B mixture-of-experts with 4B active, and a 375B flagship with 23B active parameters. IFM claims the 0.9B is the best model of its size at math, reasoning, and tool use, and that the 7B leads under 10B parameters — those are the lab's own benchmark claims, not independent results.

Independent measurement exists for the flagship. Artificial Analysis scores K2 Horizon 375B A23B at 38 on its Intelligence Index, against a median of 22 for comparable models, and lists a native 524,288-token context window. Weights are on Hugging Face and the models run on vLLM and SGLang today; hosted APIs come through partners including Cerebras, AWS, and Nebius.

The license is the part to read twice. Apache 2.0 means commercial use, modification, and redistribution with no revenue share and no usage terms bolted on later. IFM founder Eric Xing framed the release around being able to "examine, build upon, and improve the technology" — which is only true when the data ships with the weights.

Why open weights matter for your business

We made the cost case for portability in June, and the argument has not changed: an open model is leverage in a pricing conversation whether or not you ever run it. What is new here is the ladder. Six models, one training recipe, one license, spanning three orders of magnitude in size — you can prototype against the 375B through a hosted API, then move the boring 80% of your traffic to the 7B on your own hardware without changing prompt style or tokenizer behavior.

That is the practical play for most operators. You do not need a frontier model to classify support tickets, extract fields from invoices, or route inbound leads. You need a model that is cheap, predictable, and yours. Run the 7B on a machine you already pay for, keep the frontier API for the small slice of work that genuinely needs it, and measure the accuracy gap on your own data instead of on someone's benchmark table.

The published training data is the sleeper feature. If a model produces something you have to defend — to a customer, an auditor, a court — "we can show what it learned from" is a materially different position than "the vendor says it is fine."

Sources: MBZUAI, Artificial Analysis.