K2 Horizon MoVA 36B A4B
K2 Horizon MoVA 36B A4B by the MBZUAI Institute of Foundation Models is a sparse MoE model with Mixture-of-Values attention, 4B active parameters, 512K
Model Overview
Capabilities, design details, and architectural traits
K2 Horizon MoVA 36B A4B - the sparse MoVA member of the K2 Horizon fleet
K2 Horizon MoVA 36B A4B is the sparse model in the six-model K2 Horizon fleet released by the MBZUAI Institute of Foundation Models (IFM). It is a Mixture-of-Experts model that adds Mixture-of-Values attention (MoVA), applying sparsity to attention values as well as to experts. IFM positions it for cost-efficient local hosting and efficient serving.
What sets it apart
| Trait | Detail |
|---|---|
| Mixture-of-Values attention (MoVA) | Extends MoE sparsity to the computation of attention values, not just the feed-forward experts. |
| Native 512K context | 524,288-token context from the midtraining stages onward. |
| Selectable reasoning effort | Thinking depth is chosen per request via chat_template_kwargs with high, medium, and low modes; reasoning is returned in reasoning_content. |
| Multiple tool call formats | Supports json, xml, and xml_typed tool call formats, with xml as the default. |
| Fully open release | Intermediate checkpoints, training data or data-construction recipes, and training code are released under Apache 2.0. |
| One connected fleet | Shares architecture, vocabulary, training methodology, and deployment tooling with the other five K2 Horizon models. |
Open lifecycle for agent research
K2 Horizon is described by IFM as the first fully open model fleet to expose the complete development process through agentic post-training. For this model, that means researchers can study how reasoning, tool use, and agentic capabilities emerge across checkpoints rather than at a single final weight release, and adapt the released recipes to new tools and domains. GGUF versions are also published for llama.cpp, with BF16 tensors and a compatible chat template.
Serving and interfaces
The model ships with dedicated k2_horizon reasoning and tool call parsers in vLLM and SGLang, with expert parallelism for its MoE layers. Recommended inference settings are temperature=1.0 and top_p=0.95, with reasoning_effort="high" for full thinking.
Benchmark Performance
Independent evaluations · Artificial Analysis
Accuracy & Capability Details
Compare Models Side-by-Side
Evaluate specifications, pricing, and independent benchmark indices
| Model Details | |||
|---|---|---|---|
| General Info | |||
| Provider | MBZUAI Institute of Foundation Models | Anthropic | Anthropic |
| Release Date | September 3, 2026 | September 22, 2026 | September 28, 2026 |
| Knowledge Cutoff | - | - | Jun 2026 |
| Context & Limits | |||
| Context Window | 524K | 1M Best Context Window | 1M Best Context Window |
| Pricing (per 1M tokens) | |||
| Input Pricing | Free Best Input Pricing | $4 | $2 |
| Output Pricing | Free Best Output Pricing | $20 | $10 |
| Modalities | |||
| Inputs | text | textimagefile | textimagefile |
| Outputs | text | text | text |
| Benchmarks (0-100) | |||
| Intelligence Index | 25.3 | 57.6 Best Intelligence Index | 56.0 |
| Coding Index | - | - | - |
| Agentic Index | 33.5 | - | - |
Humanity's Last Exam
Extremely difficult logical reasoning and knowledge.
Long Context Reasoning
Logical reasoning over long context windows.
SciCode Benchmark
Scientific coding and mathematical modeling.
Independent evaluation data provided by Artificial Analysis. To view the latest benchmarks and full details, visit their official site.
Explore more from MBZUAI Institute of Foundation Models
Other models by MBZUAI Institute of Foundation Models
K2 Horizon 375B A23B
K2 Horizon 7B
K2 Horizon 3.7B
K2 Horizon 0.9B
Top AI Models
Leading alternatives by intelligence score
