K2 Horizon Open Model Fleet

IFM shipped six Apache 2.0 K2 Horizon models (0.9B-375B-A23B) on 3 September 2026 on Hugging Face, with day-zero support called out for vLLM, SGLang, and Ollama. It is an open fleet release, not a single chat endpoint you bolt onto a Reel pipeline.

What IFM released

The Institute of Foundation Models (IFM) announced K2 Horizon on 3 September 2026 as a connected fleet of six models:

• 375B-A23B (sparse MoE; ~23B active per token)
• 36B-A4B (MoVA sparse attention + MoE; ~4B active)
• 32B (dense)
• 7B
• 3.7B
• 0.9B

Weights and code are under Apache 2.0. Models sit under the IFM org on Hugging Face. IFM’s launch materials also emphasize opening more of the training lifecycle (intermediate checkpoints, data or data recipes, configs, logs) than a weights-only drop. Datasets follow their own licenses (for example ODC-BY where applicable).

Serving notes from the launch:

• Day-zero integration called out for vLLM, SGLang, and Ollama
• Deployment mentioned across NVIDIA, AMD, and Cerebras hardware
• Hosted API paths via IFM’s inference partners (Compass, Cerebras, Nebius) in the press materials

This post will not reprint IFM’s full benchmark tables as Versely scores. If you need the numbers, read IFM’s blog and the per-size model cards on Hugging Face and judge them against your own evals.

Why a fleet matters more than a flagship

Most open launches ship one hero checkpoint and a leaderboard screenshot. Horizon ships six sizes that share architecture family, vocabulary (with a smaller vocab on 0.9B), training methodology, and serving tooling. That design is aimed at:

1. Edge to enterprise routing. 0.9B for tightly constrained devices; 3.7B and 7B for phones and local apps; 32B / 36B-A4B for workstations; 375B-A23B for heavy enterprise serving.
2. Research that can actually reproduce stages. Intermediate checkpoints and recipes matter if you care how agentic behavior emerged, not only what the final chat model scores.
3. One mental model for ops. Same family means less glue when you scale a prototype from 7B to 32B.

IFM highlights MoVA (Mixture-of-Value Attention) on the 36B-A4B model as the efficiency play: near dense-32B class behavior with far fewer active parameters per token, per their own write-up. Treat that as vendor positioning until you reproduce it.

What this is not

K2 Horizon is not an image, video, or TTS release. It is a language / agent model fleet. Do not swap it in for a stills row, a talking-head generate, or a caption burn-in tool.

It is also not automatically a Versely catalog row. Versely meters published generate and chat routes through its own registry. An Apache 2.0 HF drop is something you can self-host with vLLM, SGLang, or Ollama, or call through IFM’s partners. Until a Horizon size is listed in Versely’s catalog, link the research and keep production creative traffic on published rows.

What I’d actually do with the news

Self-host experimenters. Pull a 3.7B or 7B build from Hugging Face, serve with Ollama or vLLM, and see whether your internal agent prompts (shot lists, routing cards, QA checklists) hold. Keep creative renders on catalog video and image models.

Platform teams. Read the Apache 2.0 terms, confirm dataset licenses separately, and decide whether Horizon becomes an internal brain behind MCP or CLI tooling. The brain that writes the brief is still not the row that spends render credits.

Everyone else. Bookmark the IFM launch post and the HF org. When you need a still or a clip, open Model Finder or browse /models and pick a published creative row. Horizon competes with open LLM fleets, not with Seedance or GPT Image 2. Soft helpers like script timing estimator and safe-area preview still sit on the creative side of the desk.

FAQ

Are all six models Apache 2.0?

IFM states models and code are Apache 2.0. Datasets use applicable licenses; read each dataset card before you redistribute training data.

Can I run Horizon in Ollama / vLLM / SGLang today?

IFM’s 3 September materials claim day-zero support for vLLM, SGLang, and Ollama. Confirm the exact model tag in the HF card and the serving project’s release notes before you promise a ship date to your team.

Is 375B-A23B “fully dense 375B”?

No. It is a sparse MoE with about 23B parameters active per token, per IFM and the HF card.

Does Versely host K2 Horizon yet?

Not claimed here. Use Hugging Face plus your own serving stack, or IFM’s partner APIs, until a catalog row exists.