K2-Horizon

K2-Horizon is a fully open decoder-only family by IFM (the LLM360 team) with a native 512K context window. The dense sizes are 0.9B, 3.7B, 7B and 32B; MoVA-36B-A4B and 375B-A23B add sparse MoE layers whose attention also routes its value projection through a set of experts (MoVA).

Axolotl trains K2-Horizon with the modeling code shipped in the checkpoints, so trust_remote_code: true is required.

Getting started

  1. Install Axolotl following the installation guide.

  2. Run the finetuning example:

    # Dense 7B, LoRA
axolotl train examples/k2-horizon/k2-horizon-7b-lora.yaml

# MoVA-36B-A4B, QLoRA + FSDP2
axolotl train examples/k2-horizon/k2-horizon-mova-36b-a4b-qlora-fsdp.yaml

Tips

  • For the MoE checkpoints, prefer an explicit lora_target_modules list over lora_target_linear: true to keep the adapter count manageable; note MoVA attention has v_experts instead of v_proj.
  • Use lora_target_modules or lora_target_linear: true to target the MoE layers; the experts are plain nn.Linear modules.
  • sample_packing: true is supported.
  • Read more on how to load your own dataset at docs.
  • The dataset format follows the OpenAI Messages format as seen here.
  • The chat template requires a thinking field on every assistant message (reasoning_content, reasoning, think, think_fast or think_faster); map yours with field_thinking, as the examples do for thinking.

Limitations

  • Cut Cross Entropy, Liger kernels and the LoRA QKV/O kernels do not cover this architecture.
  • The MoE experts are a per-expert nn.Linear loop, not the transformers v5 fused Experts layout, so the ScatterMoE / SonicMoE kernels do not apply.
  • Use FSDP rather than DeepSpeed ZeRO-3 for the MoE checkpoints: the expert loop only runs the experts a batch routes to, which can hang ZeRO-3’s per-module parameter gathering.
  • output_router_logits is off in the published configs, so no load-balancing auxiliary loss is added unless you enable it via overrides_of_model_config.