K2 Horizon
by IFM · Model · Language & code models
Report compares four model sizes trained on the same 22T tokens and describes a decoding adapter claimed to give 2.2x speedup
StatusResearchAccessOpen weightsNeeds~2 GB (a quantized build)LicenseApache-2.0
Evidence
- Sep 4IFM (MBZUAI) released six models, 0.9B to a 375B MoE, with weights, training data, checkpoints and code under Apache 2.0
- Sep 4Claimed strengths: 0.9B leads on AIME 2026 math and BFCL tool calling; 7B on SWE-bench and Terminal-Bench; 32B on GPQA Diamond
- Sep 432B translated major languages decently but produced garbled mixed-script output for low-resource languages such as Bengali, Swahili and Tamil
- Sep 3Self-reported charts show dense 32B and 36B models significantly behind Qwen3.8 27B; Gemma 4 31B not compared