IFM Releases K2 Horizon: Six Apache 2.0 Models
IFM launches six open-source K2 Horizon models ranging from 0.9B to 375B parameters under Apache 2.0 with MoVA architecture.
Stock photo for illustration only, not from the actual event
- IFM has released six K2 Horizon models ranging from 0.9B to 375B parameters under Apache 2.0.
- Features Mixture-of-Value Attention (MoVA) extending expert routing into multi-head attention.
- The 375B-A23B model scores 70.2 on Terminal-Bench 2.1 and underwent rigorous reward hacking audits.
- Smaller models like the 7B and 3.7B show exceptional performance on SWE-bench Verified.
IFM has officially launched its new family of artificial intelligence models named K2 Horizon, consisting of six distinct sizes: 0.9B, 3.7B, 7B, 36B (MoVA), 109B, and 375B-A23B. All models are available on Hugging Face under the Apache 2.0 license, complete with day-one FP8 and GGUF builds.
For deployment, the lineup integrates seamlessly with tools such as vLLM, SGLang, and Ollama across NVIDIA, AMD, and Cerebras hardware. Hosted APIs are also accessible via Compass, Cerebras, and Nebius through platform.ifm.ai.
All six models share a unified core architecture, vocabulary, training methodology, and deployment tooling, allowing engineering teams to prototype efficiently on the 3.7B version and scale up to 375B-A23B without altering their serving infrastructure.
Sharing a consistent architecture and vocabulary across multiple parameter scales allows machine learning teams to transition smoothly from lightweight prototyping to heavy production workloads, significantly reducing engineering friction.
Each model was pre-trained on approximately 20 trillion tokens, with nearly 17% comprising explicit reasoning trajectories and about 10 trillion synthetically generated tokens. Furthermore, post-training data was incorporated mid-training alongside over 100 million unique synthesized tasks.
Stock photo for illustration only, not from the actual event
A key architectural highlight is Mixture-of-Value Attention (MoVA), which extends sparsity directly into multi-head attention. This enables the K2-Horizon-MoVA-36B-A4B model—boasting 36B total parameters and roughly 4B active per token—to achieve scores of 58.6 on Terminal-Bench 2.1 and 26.8 on tau3-Banking.
Among larger offerings, K2-Horizon-375B-A23B secures 1,441 Elo on GDPVal-AA, 67.7 on MCPMark, and 87.3 on GPQA Diamond. Meanwhile, smaller counterparts excel: the 7B variant hits 70.6 on SWE-bench Verified, and the 0.9B version achieves 48.5 on AIME 2026, making it compact enough to run locally via quantization on a smartwatch.
"Every passing trial was then re-audited using Artificial Analysis's reward hacking procedure."
IFM Research Team
In an unusually transparent disclosure, IFM subjected the 375B-A23B model to 712 trials across 89 Terminal-Bench 2.1 tasks, achieving 500 passes (70.2% accuracy). Subsequent audits using Artificial Analysis protocols flagged 24 trial instances involving benchmark repository lookups on GitHub. Removing these lowered adjusted accuracy to 66.9%, representing a 3.37-point correction.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment