AuraTracer智迹闻
中文

EVENT DOSSIER

IFM Releases K2 Horizon: Six Apache 2.0 Models From 0.9B to 375B

2026-09-07 13:00 Models 🔥 42.2 heat score
1sources
1days unfolding
42.2heat score
5mentions
SummaryAI generated

The IFM laboratory under `MBZUAI` released the K2 Horizon suite on September 7, 2026, comprising six open-source models with parameters ranging from 0.9B to 375B-A23B. The suite uses the Apache 2.0 protocol and provides complete resources such as pre-trained datasets, intermediate checkpoints, and training code. All models support FP8 and GGUF formats and are available for deployment on platforms such as Hugging Face, vLLM, SGLang, and Ollama. The six models share a common core architecture and interfaces; the MoVA technology extends expert routing to the attention mechanism, enabling the 36B-A4B model to achieve better performance than dense models of the same size while maintaining parameter sparsity; the Uno adapter achieves approximately 3 times faster inference through diffusion distillation. The K2-Horizon-375B-A23B scored 70.2 on Terminal-Bench 2.1, but on SWE-bench Ver…

Related eventsRELATED EVENTS
Key entitiesKEY ENTITIES
AMDHugging FaceIFMMBZUAINVIDIA

Event frameEVENT FRAME

Launch

IFM K2 Horizon 发布六款开源大模型

Coverage · reports per dayLANGUAGE SPLIT

Entity relations
AMD × Hugging Face1AMD × IFM1AMD × MBZUAI1AMD × NVIDIA1Hugging Face × IFM1Hugging Face × MBZUAI1

SignalsSIGNALS

Keyword heat
  • IFM1
  • MBZUAI1
  • Hugging Face1
  • NVIDIA1
  • AMD1

All reports (1)SOURCES

M MarkTechPost en 2026-09-07 13:00

IFM Releases K2 Horizon: Six Apache 2.0 Models From 0.9B to 375B

The IFM laboratory under `MBZUAI` released K2 Horizon last week, offering six Apache 2.0 open-source models ranging from 0.9B to 375B-A23B. The kit includes pre-trained datasets, intermediate checkpoints, and training code, among other complete resources. All models support FP8 and GGUF formats and are available for deployment on platforms such as Hugging Face, vLLM, SGLang, and Ollama. The six models share a common core architecture and interfaces; the MoVA technology extends expert routing to the attention mechanism, enabling the 36B-A4B model to outperform dense models of the same size while maintaining parameter sparsity. The Uno adapter achieves approximately three times faster inference through diffusion distillation. K2-Horizon-375B-A23B scored 70.2 on Terminal-Bench 2.1, but it lagged behind GPT-5.6 Lu in tasks such as SWE-bench Verified.