Editorial illustration for IFM Launches K2 Horizon With Six Open-Source AI Models Up to 375B Parameters
IFM Launches K2 Horizon: Six Open-Source Models to 375B
The Institute of Foundation Models spent last week doing what most AI labs avoid: showing the whole receipt. Where a typical open-model launch means one checkpoint and a benchmark chart, IFM's K2 Horizon release covers six models ranging from 0.9B to 375B-A23B parameters, all under Apache 2.0. Alongside the weights, IFM published the pre-training corpus, intermediate checkpoints, training code, configs and detailed logs, a scope the lab is calling the largest fully open-source model launch in AI history.
IFM is the frontier lab MBZUAI launched in May 2025, and K2 Horizon is built for actual deployment rather than benchmark bragging rights. All six sizes are live on Hugging Face with FP8 and GGUF builds, day-zero support for vLLM, SGLang and Ollama, and compatibility across NVIDIA, AMD and Cerebras hardware. Hosted access runs through Compass, Cerebras and Nebius via platform.ifm.ai.
What separates this release from a simple size lineup is architectural consistency: every model shares the same core design, vocabulary and training methodology, letting teams prototype small and scale up without retooling their serving stack. The training data itself, roughly 20 trillion tokens per model, tells its own story.
The six models share a core architecture, vocabulary, training methodology, interfaces and deployment tooling. The 0.9B model uses a smaller vocabulary. That consistency is the point: teams can prototype on 3.7B and scale to 375B-A23B without changing their serving stack.
Why this matters
For developers, the appeal here isn't the 375B flagship, it's the six-way size spread plus every intermediate checkpoint and training log. That's a rare gift: you can study how a model's tool-use behavior evolves across scale instead of guessing from a single release. The JSON/XML/Markdown tool-format experiment is worth watching closely too.
If Markdown really holds an 18.5% token efficiency edge at inference while teaching semantics over syntax, that's a concrete, testable claim other labs should try to replicate, not just cite. Founders building agent stacks should note IFM is choosing Markdown as the default, a decision with real cost implications at scale.
Our skepticism is about the "100 million unique synthesized tasks" figure and the "largest fully open-source launch" framing, both of which need independent verification once researchers can actually run these checkpoints. MBZUAI backing gives IFM resources most open labs lack, so the real test is whether outside teams can reproduce the training results, not just download the weights. Watch for the first third-party benchmark runs on the 36B-A4B and 32B checkpoints.
Related Reading
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
Open Source OpenAI's GPT Transcribe Cuts Error Rate to 3.31%, Improving on GPT-4o
Open Source Study: 6.7% of Deepfake Requests on Hugging Face Targeted Children
From Archives NVIDIA Jetson Gains Nemotron 3.5 Lightning, Qwen3.8-27B for Edge AI
From Archives Startup sells service to remove safety guardrails from open-weight AI models
Common Questions Answered
What makes IFM's K2 Horizon release different from typical open-model launches?
Unlike most open-model releases that provide only one checkpoint and benchmark charts, IFM's K2 Horizon release includes six models ranging from 0.9B to 375B parameters, along with the pre-training corpus, intermediate checkpoints, training code, configs, and detailed logs. This comprehensive approach represents the largest fully open-source model launch in AI, giving developers complete transparency into the entire development process.
How does the consistent architecture across K2 Horizon models benefit developers?
All six K2 Horizon models share a core architecture, vocabulary, training methodology, interfaces, and deployment tooling, allowing teams to prototype on smaller models like the 3.7B version and scale to the 375B-A23B model without changing their serving stack. This consistency eliminates the need to redesign infrastructure or rewrite code when moving between different model sizes.
What advantage does studying intermediate checkpoints provide in the K2 Horizon release?
Access to every intermediate checkpoint and training log allows developers to study how a model's tool-use behavior evolves across different scales rather than having to guess based on a single final release. This transparency enables researchers to understand the relationship between model size and capability development in ways that weren't previously possible with typical closed releases.
What is the significance of the JSON/XML/Markdown tool-format experiment in K2 Horizon?
The Markdown tool format shows an 18.5% token efficiency edge at inference while teaching semantics over syntax, making it a concrete and testable claim worth monitoring closely. This experiment could have important implications for how models handle structured data and tool-use tasks more efficiently.
What license are the K2 Horizon models released under?
All six K2 Horizon models are released under the Apache 2.0 license, which is a permissive open-source license that allows for both commercial and non-commercial use. This licensing choice makes the models freely available for developers and organizations to use, modify, and distribute.
Further Reading
- K2 Horizon Press Release - Institute of Foundation Models
- Introducing K2 Horizon: Frontier Performance, Radically Open - Institute of Foundation Models
- IFM Releases K2 Horizon: Six Apache 2.0 Models From 0.9B to 375B - MarkTechPost
- IFM Releases K2 Horizon, Six Open Models From 0.9B to 375B With the Training Data - Pasquale Pillitteri
- IFM Releases K2 Horizon: Six Open Models With Data, Code, and ... - NYU Shanghai Research & Innovation
