K2 Horizon
#26 in Open-Source-SpraakmodelleMbzuai · siet 3. September 2026 · 4× · tolest 05. Sept. 2026
37
Momentum
K2 Horizon is a family of six open language models, released on September 3, 2026 by the Institute of Foundation Models (IFM) at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), ranging from 0.9 to 375 billion parameters. All models are released under the Apache 2.0 license with weights, training code and (largely) training data, covering deployment scenarios from wearables to enterprise servers. The flagship 375B-A23B is a Mixture-of-Experts model with 23B active parameters, native 512K-token context, and is text-only (no image/video input). Models are available via Hugging Face, vLLM, SGLang, and APIs from Cerebras, AWS, Nebius and Compass.
Momentum-Verloop
08.06.06.09.
Features
| Key Benchmark (%) | Flagship 375B-A23B: 70.2% on Terminal-Bench 2.1; 42.6% on SWE-Bench (per IFM figures) |
| Context Window (Tokens) | 512K (native 524,288 tokens, from midtraining stages onward) |
| License | Apache 2.0 (models and code) |
| Multimodality | Text-only (input and output), no image/video support |
| Platform | Hugging Face, vLLM, SGLang, Ollama; API via Compass, Cerebras, AWS and Nebius |
| Price per 1M Tokens | 375B-A23B: $0.00 (no official API pricing published yet, per Artificial Analysis) |
| Release Date | September 3, 2026 |