MBZUAI releases six K2 Horizon models from 0.9B to 375B parameters

• K2 Horizon is an Apache-2.0 fleet spanning 0.9B to 375B parameters, with all six sizes listed in testingcatalog's release summary.
• The release includes final weights plus checkpoints, code, data recipes, logs, and evaluations, as kimmonismus's post noted.
• vLLM has day-zero support, and vllm_project's announcement says the family includes reasoning and tool calling with 512K context from 3.7B upward.
• The 375B-A23B flagship scored 47 on Artificial Analysis's Intelligence Index, a 30-point gain over K2 Think V2, according to ArtificialAnlys's benchmark thread.
• MoVA moves expert routing into attention, while Uno is a diffusion adapter for parallel token blocks, as testingcatalog's explainer described.

IFM's technical launch post puts the 0.9B model on watches and glasses, and the 3.7B and 7B variants on phones and other on-device applications. It also says the family was pretrained on roughly 20 trillion tokens, including nearly 17% explicit reasoning trajectories and about 10 trillion synthetic tokens.

IFM released artifacts from pretraining through reasoning and agentic post-training, not only the final checkpoints. Its release notes specify that where data cannot be redistributed, the bundle substitutes source descriptions, filtering and construction methods, and mixture composition.

The published record covers:

• Final model weights and evaluation results.
• Intermediate checkpoints across training.
• Training code, configurations, and recipes.
• Training data or data-construction recipes, plus mixture compositions.
• Fine-grained training logs.
• Architecture details and post-training branches for reasoning, instruction following, coding, tool use, and agentic work.

The six models share core architecture, training methodology, interfaces, evaluation infrastructure, and deployment tooling, according to IFM's launch post. The 0.9B model alone uses a smaller vocabulary.

• 0.9B: constrained edge devices, including watches and glasses.
• 3.7B and 7B: phones and other on-device applications.
• 32B: the largest dense model, aimed at local workstations and efficient serving.
• 36B-A4B: a sparse MoVA model for the same local-serving range.
• 375B-A23B: the largest sparse model, intended for demanding enterprise workloads.

All six include quantization support, IFM says.

MoVA applies sparsity inside multi-head attention, extending the expert routing conventionally used in feed-forward layers. testingcatalog's explainer identifies it as one of two architecture pieces released with Horizon.

• MoVA: IFM says it remains compatible with FlashAttention, grouped-query attention, and sparse attention. The 36B-A4B model has 36B total parameters and activates about 4B per token; IFM says it performs only slightly below the dense 32B model under the same training conditions.
• Uno: a LoRA-style diffusion adapter that leaves the autoregressive Horizon parameters frozen and learns to generate token blocks in parallel. IFM characterizes the result as a lossless inference speedup in its technical description.

Artificial Analysis places K2 Horizon 375B-A23B at 47 on its Intelligence Index, up from K2 Think V2's 17. The model swaps its predecessor's 70B dense design for a 375B total, 23B-active MoE and doubles context from 262K to 512K tokens.

• On GDPval-AA, it recorded 1,430 Elo versus MiniMax-M3's 1,380; on τ³-Banking, it scored 34.2% versus 15.3%, per ArtificialAnlys's comparison.
• It trailed MiniMax-M3 on GPQA Diamond, 87.3% versus 92.9%, and Humanity's Last Exam, 32.0% versus 39.0%, in ArtificialAnlys's results.
• Its 26% AA-Omniscience hallucination rate improved on K2 Think V2's 71%, but ArtificialAnlys's analysis attributes much of that result to abstention: Horizon attempted 40% of questions, declined 60%, and retained 18% accuracy.

The published vLLM recipe for the 36B-A4B variant sets --max-model-len 131072, even as vllm_project's announcement describes 512K context for models from 3.7B upward.

That recipe lists 37.44B stored parameters including embeddings and 5.95B active parameters per token, and names H100, H200, B200, GB200, and AMD MI300X-class hardware among its supported accelerators.