---
title: "K2 Horizon: What a Fully Open Model Fleet Changes"
slug: "k2-horizon-fully-open-model-fleet-enterprise"
author: "ibl.ai Engineering"
date: "2026-09-07 11:00:00"
category: "Premium"
topics: "open source AI, open weights, K2 Horizon, MBZUAI, foundation models, model-agnostic, self-hosted AI"
summary: "MBZUAI's Institute of Foundation Models released six Apache-2.0 models from 0.9B to 375B parameters on one day — with training code, data mixtures, intermediate checkpoints and evaluation logs. For enterprises the shared architecture matters more than any single model."
banner: ""
thumbnail: ""
linkedin: |
  On 3 September, MBZUAI's Institute of Foundation Models released K2 Horizon: six foundation models from 0.9B to 375B parameters, all Apache-2.0, all on the same day.

  The sizes are 0.9B, 3.7B, 7B, 32B, 36B-A4B and 375B-A23B. Each was pretrained on roughly 20 trillion tokens.

  What makes this different from a normal open-weights drop is what ships alongside the weights: training code and data mixtures. One caveat the headline glosses — the 3.7B and 7B are complete today, while the 375B and 36B model cards say checkpoints, data and code "will be released." Most "open" models give you a weights file and a licence. This gives you the recipe, for most of the fleet.

  But the detail enterprises should actually care about is quieter than the 375B headline.

  Every size in the fleet shares the same architecture, vocabulary, training recipe, interfaces and evaluation tooling. That means a team can route work by size — a 0.9B model on a device, a 7B in a branch office, the 375B in the datacenter — without rewriting integrations, re-tuning prompts or maintaining separate evaluation harnesses for each.

  That is the part that changes an architecture rather than a benchmark. Most cost optimization in production AI is really routing: send the trivial classification to something small and local, reserve the expensive model for work that needs it. Routing across models from different labs means reconciling different tokenizers, different tool-calling conventions and different failure modes. Routing within one fleet does not.

  Apache-2.0 with published data also clears procurement questions that block regulated deployments — what the model was trained on, whether it can be run inside a perimeter, whether it can be fine-tuned without a vendor's permission.

  With ibl.ai you own all the code and the data — self-hosted inside your own perimeter, model-agnostic across any LLM, usage-based with no per-seat pricing, deployable anywhere from your own cloud to a fully air-gapped network. An open fleet like this is only useful if your platform can actually run it.

  #iblai #OpenSource #EnterpriseAI #OpenWeights #AgenticAI #SelfHosted
---

## The Short Answer

**On 3 September 2026, MBZUAI's Institute of Foundation Models released K2 Horizon — six Apache-2.0 models from 0.9B to 375B parameters, shipped with training code and data mixtures for the smaller sizes, the largest models' artifacts promised. The enterprise significance is not the 375B model; it is that every size shares one architecture, so work can be routed by size without rewriting anything. With ibl.ai you own all the code and the data, model-agnostic across any LLM.**

An open-weights release you cannot deploy inside your own perimeter is a press release. The reason this one matters to an enterprise is structural, and it is not the headline parameter count.

## What exactly did MBZUAI release in K2 Horizon?

Six foundation models, all at once, all under Apache-2.0: **0.9B, 3.7B, 7B, 32B, 36B-A4B and 375B-A23B**. Each was pretrained on approximately **20 trillion tokens**.

The Institute of Foundation Models — the research lab MBZUAI launched in May 2025 — [published the fleet on 3 September 2026](https://ifm.ai/k2/).

The release goes past the weights — though not uniformly, and the distinction matters. The **3.7B and 7B ship complete today** with data, recipe and training code.

The model cards for the **375B-A23B and 36B-A4B state that checkpoints, data and training code "will be released"**, and the 32B is currently a Stage-1 checkpoint. "Fully open" is the direction of travel; today it is fully true of the smaller models.

Even partially, that goes well beyond the norm: most models described as "open" publish a weights file under a bespoke licence, which lets you run the model but not reproduce, audit or genuinely extend it.

The smaller models are not filler. The 0.9B, 3.7B and 7B set state of the art at their respective scales on coding and agentic benchmarks, which is what makes the fleet framing credible rather than a marketing device wrapped around one flagship.

## Why does a shared architecture across model sizes matter more than the largest model?

Because production AI cost is mostly a routing problem, and routing is where heterogeneous models get expensive.

A realistic deployment does not send every request to the biggest model available. It classifies, extracts and summarizes with something small, and reserves the expensive model for work that genuinely needs it.

That is where most of the savings in a production system come from — not from negotiating a better per-token rate.

The obstacle is that routing between models from different labs means reconciling different tokenizers, different tool-calling conventions, different context limits and different failure modes.

Each of those is integration work, and each is a place your evaluation harness stops being comparable.

K2 Horizon's fleet shares architecture, vocabulary, training recipe, interfaces and evaluation tooling across all six sizes. A team can move between them, or route by size, without rewriting integrations or maintaining a separate harness per model.

**A 0.9B model on a device and a 375B model in a rack that answer to the same interface are a different engineering proposition from six unrelated models that happen to be free.**

## What does an Apache-2.0 licence with published training data change for regulated buyers?

It answers the questions that usually stall a deployment before the technical evaluation even starts.

Regulated procurement asks three things about a model that vendor-hosted APIs answer badly: what was it trained on, can we run it entirely inside our own perimeter, and can we fine-tune it without asking permission.

A permissive licence with published data mixtures answers all three in the affirmative, in writing, before legal review — for the sizes where the data has actually shipped, which today means the smaller models rather than the 375B.

This is the practical difference between "open weights" and "fully open". A model you can download but whose training data is undisclosed still leaves a bank or an agency unable to answer a supervisor's question about provenance.

Published data mixtures and evaluation logs turn that from an unknown into a document.

## Can an enterprise actually run a 375B model on its own infrastructure?

Most cannot, and that is exactly why the fleet structure is the story.

A 375B-A23B mixture-of-experts model is a serious infrastructure commitment.

The realistic pattern for most institutions is to self-host the small and mid-sized models where privacy and latency matter, and reach for a hosted frontier model only for the workloads that need it — with the option to bring more in-house as hardware economics move.

That hybrid is only coherent if the platform treats models as interchangeable.

If your authentication, retrieval, guardrails and audit logging are welded to one provider's API, you cannot mix a self-hosted 7B with a hosted frontier model behind one policy — and the open fleet is a curiosity rather than an option.

We have written before about why [vendor lock-in is the real risk in model-agnostic AI](/blog/model-agnostic-ai-the-real-risk-is-vendor-lock-in), and an open fleet is the clearest case of a benefit you forfeit entirely without the architecture to use it.

## How does ibl.ai use an open model fleet like this?

By treating the model layer as the replaceable part, which is what makes any of the above available to you.

With ibl.ai you own all the code and the data. The platform is deployed on your own infrastructure with full source code access, runs any LLM — hosted frontier models, self-hosted open-weight models, or both behind one routing policy — and is priced by usage rather than per seat.

It deploys to your own cloud, on-premise, GovCloud, or a fully air-gapped network where no external API is reachable at all.

For an air-gapped deployment, a fully open fleet is not a cost optimization but a precondition: there is no hosted model to call, so the models you can run are exactly the models you can hold.

A release that spans device-scale to datacenter-scale under one architecture makes that deployment considerably more practical than it was before 3 September.

ibl.ai is family-owned and operated from New York, NY — a long-term partner rather than a vendor selling licences, which for institutions building on infrastructure they intend to own for a decade is not a small consideration.

*Related reading: [what published inference benchmarks reveal about the cost floor](/blog/inference-cost-floor-jalapeno-published-benchmarks), and [GPT-6 Astra and the harness footnote](/blog/gpt-6-astra-arc-agi-3-model-agnostic-architecture) on why leaderboard comparisons deserve scrutiny.*

*Source: model sizes, licence, training-token counts and the contents of the release are from [the Institute of Foundation Models' K2 Horizon announcement](https://ifm.ai/k2/) and its [press release](https://ifm.ai/k2/press-release/).*

*Related: [digital sovereignty: why agencies need model-agnostic AI](/blog/government-ai-digital-sovereignty-model-agnostic) — why an Apache-2.0 fleet changes what is buildable inside an air-gapped enclave.*

## Why does owning the AI stack matter?

**ibl.ai is the agentic AI platform where you own all the code and the data. You self-host the entire stack inside your own perimeter, run it model-agnostic across any LLM and switch anytime, and pay by usage with no per-seat pricing — so you can deploy anywhere: your cloud, on-premise, GovCloud, or fully air-gapped.**

- **You own all the code and the data.** Full source code under a perpetual license, running on your infrastructure. Not API access to someone else's platform — the stack itself is yours.
- **Model-agnostic.** Run any LLM — Claude, GPT, Gemini, Llama, Command, or your own fine-tune — and switch providers without rewriting the platform.
- **No per-seat pricing.** Usage-based billing against a budget cap you set. Cost tracks what your organization actually uses, not how many people you employ.
- **Deploy anywhere.** Your cloud, your VPC, on-premise, GovCloud, or a fully air-gapped network with no outbound connectivity.

1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University.

ibl.ai is family-owned and operated from New York, NY — a U.S.-headquartered, domestically-owned long-term partner, not a vendor that sells licenses and moves on.