Vendor · Language models
Institute of Foundation Models has 6 models in the directory: 6 language models.
In the directory: 6 language models. 6 of them with open weights that can run locally.
Every model there ever was. Models with a profile lead to every benchmark, price and source.
| K2 Horizon 0.9B | Sep 3, 2026 | Open weights | n/a | n/a | n/a | 131K | 44 |
| K2 Horizon 3.7B | Sep 3, 2026 | Open weights | n/a | n/a | n/a | 524K | 47 |
| K2 Horizon 32B | Sep 3, 2026 | Open weights | n/a | n/a | n/a | 524K | 41 |
| K2 Horizon 375B-A23B | Sep 3, 2026 | Open weights | n/a | n/a | 929 | 524K | 45 |
| K2 Horizon 7B | Sep 3, 2026 | Open weights | n/a | n/a | n/a | 524K | 53 |
| K2 Horizon MoVA 36B-A4B | Sep 3, 2026 | Open weights | n/a | n/a | n/a | 524K | 45 |
Language model · Sep 3, 2026
K2 Horizon 0.9B is the family’s compact dense model. It supports up to 131,072 tokens with YaRN RoPE scaling and uses an additional on-policy distillation stage, which IFM calls MOPD, after pretraining, midtraining, and RL.
Language model · Sep 3, 2026
K2 Horizon 3.7B brings the dense architecture to a smaller scale while retaining a native context of 524,288 tokens. Its published training path uses separate math, code, and STEM-code RL experts, merges their weights, and then applies FTPO as a final training stage.
Language model · Sep 3, 2026
K2 Horizon 32B is the family’s large dense model, handling text in a native context of 524,288 tokens. Its documented training runs through pretraining, four midtraining stages, and two SFT phases, with the corresponding checkpoints released publicly.
Language model · Sep 3, 2026
K2 Horizon 375B-A23B is the largest model in IFM’s open fleet. Its MoE architecture combines 375 billion total parameters and activates 23 billion per token. IFM also releases intermediate checkpoints from pretraining, midtraining, reinforcement learning, and SFT, making several stages of its development available for study.
Language model · Sep 3, 2026
K2 Horizon 7B is a dense model with a 512K context for coding, reasoning, and multi-step tasks. After reinforcement learning, it receives an extra FTPO stage that the model card says is intended to reduce repetitive loops in long generations. The smaller 3.7B model follows a similar training path.
Language model · Sep 3, 2026
K2 Horizon MoVA 36B-A4B complements the dense 32B model with a sparse architecture. Mixture-of-Value Attention routes attention through experts, while roughly 4 billion of its 36 billion parameters are active per token.
| Vendor | Models | Model types | Best language model | Top 10 | Open weights | New in 2026 |
|---|---|---|---|---|---|---|
| Anthropic | 33 | 33 language models | Claude Opus 5.5 (Rank 1 to 3) | 4 | 0 | 13 |
| OpenAI | 63 | 46 language models, 4 image models, 2 video models, 9 audio models, 2 specialist models | GPT-6 Astra (Rank 4 to 8) | 2 | 4 | 27 |
| 66 | 43 language models, 4 image models, 6 video models, 11 audio models, 2 specialist models | Gemini 4 Carbon (Rank 1 to 5) | 2 | 19 | 26 | |
| Meta | 32 | 29 language models, 1 image model, 1 video model, 1 specialist model | Muse Wonder Full (Rank 6 to 9) | 2 | 26 | 7 |
| DeepSeek | 23 | 23 language models | DeepSeek V4.1 Flash 1017 (Rank 13 to 20) | 0 | 23 | 9 |
| Moonshot AI | 12 | 12 language models | Kimi K3.1 Pro (Rank 10 to 13) | 0 | 10 | 7 |
| xAI | 16 | 12 language models, 2 image models, 2 video models | Grok 4.7 (Rank 18 to 24) | 0 | 1 | 12 |
| Institute of Foundation Models | 6 | 6 language models | not ranked yet | 0 | 6 | 6 |
By release date. Grey: retired models.
6 releases since January 2026, by release date.
The directory lists 6 K2 Horizon models from Institute of Foundation Models: 6 language models. It counts every model Institute of Foundation Models ever released, including retired ones. 6 are currently available, the newest being K2 Horizon 0.9B, K2 Horizon 3.7B and K2 Horizon 32B.
K2 Horizon 0.9B, released on Sep 3, 2026. The timeline below shows every release by date.
Yes, 6 of the 6 models have open weights. Each model profile lists how much memory it needs.
K2 Horizon 375B-A23B with 929 tokens per second according to Artificial Analysis. A typical model manages 98.
Up to 524,288 tokens (≈ 790 A4 pages), which 5 of the current models reach, including K2 Horizon 3.7B. Every model's value is in the model list.
Data as of Nov 1, 2026