Meet K2 Horizon.

Publishing the training process alongside the models is what makes this release genuinely useful for builders.

I like that every model size has a clear purpose. Choosing based on the actual workload makes far more sense than defaulting to the largest option.

The range from 0.9B to 375B opens up a lot of deployment possibilities.... I’m curious to see what developers build across both ends of the lineup.

@shauryr Congratulations on the team!

Open weights are great, but opening the training process is what really moves research forward.

The fact that the whole family shares interfaces, tools and a connected architecture also has a lot of practical value.

wrong timing bro

For companies, this structure can be especially useful.

The 7B numbers look insane

For companies building core products on AI, that distinction between access and ownership matters a lot.

K2 Horizon is making a broader argument about open AI: access to the model matters, but access to the process behind the model may matter even more

This is the kind of release where the artifacts matter as much as the benchmark numbers

The moment a smaller model clears the task reliably, using the biggest one becomes a pretty expensive default.

Transparency this rare should be the standard.

“Don’t trust, verify” is a much better standard for AI releases.

Would love to see more labs be this transparent.

IFM just raised the bar

The “right-size the model for the job” approach makes a lot more sense for production

I really like the full-stack openness here.

Great work by the community!

That's great. Every model ha different size

Having access to the path, not just the final destination, changes what researchers can study

Fleet math under “Meet K2 Horizon”: six sizes 0.9B→375B-A23B, Apache 2.0 weights+code, data or recipes, checkpoints. Flagship activates ~23B/token. IFM self-cut Terminal-Bench 70.2%→66.9% after reward-hack audit. IFM blog 3 Sep.

This is great 😃 This is how transparency should be

LFG !!!

The full training record is what really separates K2 Horizon from a typical open weights release.

The combination of performance and openness here is genuinely exciting.

Private deployment gets a lot more interesting when the model family spans this many sizes.

@IFM_AI Let's Collab

Open evaluations combined with an open training record create a much stronger foundation for independent research and reproducibility.

K2 Horizon looks like a seriously exciting open release.

This is what “open” should actually feel like.

congrat...chat pege?

the real value here is opening the recipe, not just releasing the weights.

More labs publishing failure modes this openly would make model evaluation a lot healthier