What's this about?
IFM released K2 Horizon: six connected models from 0.9B to 375B-A23B parameters, Apache 2.0, with intermediate checkpoints, training data or recipes, code, configs and fine-grained logs — billed as the first open family to expose the entire pipeline through agentic post-training.
What exactly does "fully open" mean beyond open weights?
Most "open" models ship final weights only — the cake, not the recipe. K2 ships the recipe: what data went in, in what order, with what settings, and checkpoints along the way. That's what researchers need to study *how* capabilities form, and it's the standard almost nobody meets.
What's the community saying?
HN 264 points, 84 comments
- Strong support for the principle: fully open models — data, code, process — are the only way to know what's underneath what you're using.
- Launch-day sloppiness undercut the message: the announcement link hit a login wall ("Radically Open, behind http login"), and two training repos were empty at launch.
- Benchmark skepticism: by IFM's own charts the dense 32B trails Qwen 3.8 27B, and the comparison sets change between tasks.
- Early hands-on with the small model found it weak at coding — hallucinating APIs when challenged.
- Naming complaint: "K2" collides with Kimi K2, of all the available names.