K2 Horizon is a six-model open family from the Institute of Foundation Models at MBZUAI. It ranges from a 0.9B dense model for constrained devices to a 375B mixture-of-experts model with about 23B active parameters per token.
The family is more interesting than a single flagship release because it gives developers a shared model line across edge, laptop, workstation and server deployments. Final weights and model code are released under Apache 2.0. The project’s broader transparency package needs careful wording: IFM commits to training data, code, recipes and intermediate artifacts, while individual model repositories still say some of those materials will be released.
This is research-based coverage of IFM’s official release, technical overview and 375B-A23B model card. We have not run the models or independently reproduced project benchmarks.
The six K2 Horizon models
| Variant | Architecture | Active parameters | Practical deployment tier |
|---|---|---|---|
| K2 Horizon 0.9B | Dense | 0.9B | Mobile, edge and constrained devices |
| K2 Horizon 3.7B | Dense | 3.7B | Laptop and compact local workloads |
| K2 Horizon 7B | Dense | 7B | Local workstation and single-GPU use |
| K2 Horizon 32B | Dense | 32B | High-memory workstation or server |
| K2 Horizon 36B-A4B | MoE | About 4B per token | Efficient server/local MoE workloads |
| K2 Horizon 375B-A23B | MoE | About 23B per token | Multi-GPU production inference |
Active parameters describe compute routed for each token, not the complete memory footprint. A 375B-A23B checkpoint still requires storage and memory for the full expert weights.
What “open” means here
IFM releases the final model weights and model code under the permissive Apache License 2.0. That supports commercial use, modification and redistribution subject to the licence terms.
The release also promises open training data, training and evaluation code, recipes, checkpoints and methodology. However, the current 375B repository explicitly says intermediate checkpoints, training data and training code will be released. The accurate present-tense conclusion is therefore:
- final weights and inference/model code are available;
- the wider reproducibility package is announced and partially documented;
- not every promised training artifact should be described as already downloadable.
Context and output
The 375B-A23B model card documents a 524,288-token context window. Its example vLLM deployment on eight H200 GPUs uses --max-model-len 131072, a practical serving configuration rather than proof that the architecture is limited to 131K.
Context, quantization and concurrency all consume memory. Choose a context limit based on the workload instead of configuring the maximum for every request.
Deployment support
IFM documents day-zero support across vLLM, SGLang and Ollama, with hardware work involving NVIDIA, AMD and Cerebras. The open weights also make dedicated or self-hosted deployments possible without depending on one model API.
The release names Compass, Cerebras and Nebius as API access partners. We did not find a single clear, primary-source public price that could be treated as canonical across those providers. Hosted availability therefore should not be confused with one universal K2 API price.
For local deployment, start with the 0.9B, 3.7B or 7B dense checkpoints. The 36B-A4B MoE can offer lower active compute than its total parameter count suggests, but the full weights still need to fit. The 375B-A23B model is a server-class deployment.
Which K2 Horizon size should you choose?
| Need | Start with | Why |
|---|---|---|
| Edge or on-device experiments | 0.9B | Smallest memory and compute target |
| Laptop assistant or extraction | 3.7B | Better capability while remaining compact |
| Local development quality | 7B | Familiar single-GPU/local tier |
| Dense high-quality deployment | 32B | More capacity without MoE serving complexity |
| Efficient MoE serving | 36B-A4B | About 4B active per token |
| Maximum family capability | 375B-A23B | Flagship, but requires serious infrastructure |
Run the smallest model that meets your evaluation threshold. Our local AI guide covers memory, quantization and context tradeoffs, while AI deployment and hosting helps choose between local, dedicated and managed inference.
Benchmarks require attribution
IFM reports competitive reasoning, coding and multilingual results across the family. Those numbers are project-reported and depend on prompts, harnesses, inference settings and judges. They are useful for shortlisting, not proof that K2 universally beats another model.
Build evaluations from your own tasks: structured output, tool use, code repair, latency, memory use and cost per successful result. See AI Testing & Evaluation for a production-oriented framework.
API versus self-hosting
Use a partner API when you want to evaluate quickly without operating hardware. Verify the provider’s exact model ID, context, region, retention policy and price because those can differ from the base model card.
Self-host when data control, customization or predictable dedicated capacity matters enough to justify operations work. The Apache 2.0 licence makes that commercially practical, but it does not eliminate infrastructure, security and evaluation costs.
Verdict
K2 Horizon’s strength is breadth: one Apache-2.0 family spans edge experiments through multi-GPU inference. The final weights are genuinely available, while parts of the promised training-transparency package remain forthcoming. Developers should choose by deployment envelope and measured task quality, not by assuming that “23B active” makes the 375B checkpoint a small local model.