K2-Horizon

Training data disclosure: Partial list

Level determined for 0% of benchmark cells.

• TestTest data seen0
• SeenSeen, split unclear0
• TrainTraining data seen0
• Rel.Related data seen0
• NoNot seen0
• ?Unknown15

Other declared training data

• IFM/TxT360-v2 (broad web data: may contain benchmark test sets) · split: unknown

  Subsets web-high-nltk-qa, web-high-medium, txt360-qa (web and QA text); card states information cutoff Q4 2024.

  Evidence

  “Data | [Hugging Face](https://huggingface.co/datasets/IFM/TxT360-v2) | Available”

  “Pre-training sources for the K2 Horizon training data release.”

• IFM/K2-Horizon-Pretrain-Data (not yet matched to a dataset) · split: unknown

  Listed in model card metadata but the HF dataset repo is not publicly accessible (HTTP 401 on 2026-10-15).

  Evidence

  “IFM/K2-Horizon-Pretrain-Data”

• IFM/K2-Horizon-Midtrain-Data (not yet matched to a dataset) · split: unknown

  Listed in model card metadata but not publicly accessible (HTTP 401 on 2026-10-15). Mid-training recorded as pretraining stage.

  Evidence

  “IFM/K2-Horizon-Midtrain-Data”

• IFM/Pretrain-Behaviors (not yet matched to a dataset) · split: unknown

  Part of the K2 Horizon dataset series; synthetic behaviour data. Which stage/model uses it is not stated.

  Evidence

  “Behavior-focused text covering reasoning, planning, data science, games, general content, and format rewriting.”

• IFM/Math-Reasoning (not yet matched to a dataset) · split: unknown

  K2 Horizon dataset series; stage not stated (recorded as pretraining by default). Problem sources not listed.

  Evidence

  “Mathematical problem-solving, rewriting, and dialogue data for reasoning-oriented language-model training.”

• IFM/Code-Reasoning (not yet matched to a dataset) · split: unknown

  K2 Horizon dataset series; stage not stated (recorded as pretraining by default). Problem sources not listed.

  Evidence

  “Code problem-solving data with reasoning, direct-answer, and task-synthesis subsets.”

• IFM/SFT-Reasoning (not yet matched to a dataset) · split: unknown · fine-tuning

  Subsets instruction-following and 3efforts-pretrain.

  Evidence

  “Instruction-following and reasoning data prepared for supervised fine-tuning.”

• RL data (expert models: math, code, STEM, IF, search, tool use, knowledge work) (not yet matched to a dataset) · split: unknown · fine-tuning

  RL prompt data not described or listed in any card.

  Evidence

  “We trained five expert models from the final checkpoint of Midtraining Stage 4: knowledge work, IF, search, tool use, and reasoning. We then merged the expert models.”

Benchmark cells

Classic LLM benchmarks

Recent LLM benchmarks

Notes

One entry for the family: all six cards point to the same data release (IFM/TxT360-v2 and the K2 Horizon dataset series), but per-size recipes differ: 3.7B/7B/32B/MoVA cards list 22.9T pre-training tokens, 375B-A23B lists two phases (7.08T + 8.05T), 0.9B lists 5T plus multi-teacher distillation; post-training differs (SFT only for 32B and MoVA; RL experts + merge + SFT for 3.7B/7B/375B; RL + on-policy distillation for 0.9B). No technical report yet (cards dated 2026-09-28 say end of September 2026). The ifm.ai blog returned HTTP 403 to automated fetches, so nothing is quoted from it.

Sources

• IFM/K2-Horizon-32B model card · model card · retrieved 2026-10-15
• IFM/K2-Horizon-0.9B model card · model card · retrieved 2026-10-15
• IFM/K2-Horizon-3.7B model card · model card · retrieved 2026-10-15
• IFM/K2-Horizon-7B model card · model card · retrieved 2026-10-15
• IFM/K2-Horizon-MoVA-36B-A4B model card · model card · retrieved 2026-10-15
• IFM/K2-Horizon-375B-A23B model card · model card · retrieved 2026-10-15
• IFM/TxT360-v2 · dataset card · retrieved 2026-10-15
• IFM/SFT-Reasoning · dataset card · retrieved 2026-10-15
• IFM/Math-Reasoning · dataset card · retrieved 2026-10-15
• IFM/Code-Reasoning · dataset card · retrieved 2026-10-15
• IFM/Pretrain-Behaviors · dataset card · retrieved 2026-10-15