The Institute of Foundation Models has published K2 Horizon, a set of six interconnected models ranging from around 0.9 to 375 billion parameters, all under the Apache 2.0 licence. The novelty isn't the size of the largest one: it's the release of a complete range rather than a single model.
Every week we cover open model releases. This one deserves attention for its form rather than its scores.
The problem it solves
When building a real product, you almost never use just one model. We've been saying this for months: standard practice is to route — send simple tasks to a cost-effective model and only scale up when necessary.
But routing assumes you have several compatible models available. Assembling a chain with a small model from one vendor, a mid-sized one from another and a large one from a third creates practical difficulties: formats differ, behaviours vary, instructions don't carry over.
A coherent range removes that friction. The models share an architecture, a tokenizer, a way of following instructions. What works on the small one works on the large one, and scaling up becomes a parameter change rather than a rewrite.
What each size enables
Around one billion. Runs on a phone or a connected device. Useful for classification, intent detection, filtering. It doesn't reason, and it doesn't need to.
A few billion. Runs on an ordinary laptop. Suitable for information extraction, reformulation, simple document processing.
A few tens of billions. This is the segment we described as decisive regarding Qwen 3.8 at 27 billion: a single graphics card, serious capabilities, full confidentiality.
Several hundred billion. Requires server infrastructure. This is the level where you look for complex reasoning and long-running agentic work.
Apache 2.0 across the entire range is a significant choice. As we explained in our article on licences: this licence permits commercial use without strings attached and includes an explicit clause on patents, which many legal departments require. An entire range under a permissive licence means you can build a complete architecture without negotiating with anyone.
The caveats
No independent evaluation at this stage. The announced performance figures are the lab's own, and a coherent range doesn't guarantee that each size is competitive against the specialised models in its category.
Coherence has a cost. A model designed to belong to a family sometimes makes architectural compromises that a standalone optimised model wouldn't.
Small remains small. A one-billion-parameter model has real limits, and belonging to a nice range doesn't remove them.
What this says about the moment
This release illustrates a shift we've been documenting for weeks. The competition is no longer mainly about the most powerful model: it's about what makes a model usable.
We've seen it with agent harnesses, with the standardisation of protocols, with cache optimisation. Releasing a coherent range follows the same logic: you're no longer selling a capability, you're delivering infrastructure.
What to take away
For someone building things, this form of release is often worth more than a single, more capable model.
The reason is the one we keep repeating in all its forms: what matters isn't the best choice today, it's the ability to change tomorrow. A coherent range under a permissive licence lets you start small, scale up on tasks that demand it, and scale back down when progress makes the small model sufficient.
It's less spectacular than a record. It's what remains once the records have been beaten by someone else.