---
base_model: IFM/K2-Horizon-MoVA-36B-A4B
base_model_relation: quantized
license: apache-2.0
language:
  - en
library_name: gguf
pipeline_tag: text-generation
tags:
  - gguf
  - llama.cpp
  - k2-horizon
  - long-context
  - 512k-context
  - mixture-of-experts
  - mova
---

> [!IMPORTANT]
> **Compatibility:** These GGUF files require a `llama.cpp` build with K2 Horizon architecture support. Until upstream support lands, use the [MBZUAI-IFM fork](https://github.com/MBZUAI-IFM/llama.cpp/tree/model/K2Horizon).

# K2-Horizon-MoVA-36B-A4B GGUF

GGUF quantization of [IFM/K2-Horizon-MoVA-36B-A4B](https://huggingface.co/IFM/K2-Horizon-MoVA-36B-A4B), a sparse Mixture-of-Experts model with Mixture-of-Values attention (MoVA), 36B total parameters, and approximately 4B active parameters per token. The source checkpoint supports a native context length of **524,288 tokens (512K)**.

## Benchmarks

![K2-Horizon-MoVA-36B-A4B benchmark results](assets/k2-horizon-mova-36b-a4b-benchmarks.png)

*Benchmark results reported by IFM for the original K2-Horizon-MoVA-36B-A4B checkpoint.*

## GGUF files

| Quantization | File | Size |
| --- | --- | ---: |
| Q3_K_M | [K2-Horizon-MoVA-36B-A4B-Q3_K_M.gguf](K2-Horizon-MoVA-36B-A4B-Q3_K_M.gguf) | 17.66 GB |
| Q4_K_M | [K2-Horizon-MoVA-36B-A4B-Q4_K_M.gguf](K2-Horizon-MoVA-36B-A4B-Q4_K_M.gguf) | 22.37 GB |
| Q4_0 | [K2-Horizon-MoVA-36B-A4B-Q4_0.gguf](K2-Horizon-MoVA-36B-A4B-Q4_0.gguf) | 21.28 GB |
| Q5_K_M | [K2-Horizon-MoVA-36B-A4B-Q5_K_M.gguf](K2-Horizon-MoVA-36B-A4B-Q5_K_M.gguf) | 26.44 GB |
| Q6_K | [K2-Horizon-MoVA-36B-A4B-Q6_K.gguf](K2-Horizon-MoVA-36B-A4B-Q6_K.gguf) | 30.77 GB |
| Q8_0 | [K2-Horizon-MoVA-36B-A4B-Q8_0.gguf](K2-Horizon-MoVA-36B-A4B-Q8_0.gguf) | 39.83 GB |

The model is text-only; no vision projector is required. SHA-256 checksums are provided in [`SHA256SUMS.txt`](SHA256SUMS.txt).

## Chat template

Each GGUF embeds the llama.cpp-compatible chat template. [`chat_template.jinja`](chat_template.jinja) is a matching external copy for tools that require one. The original source template is retained as [`chat_template.upstream.jinja`](chat_template.upstream.jinja) for runtimes with full Jinja support.

`xml` is the default tool-call format. Use `--chat-template-kwargs` to select `json` or `xml_typed` when required.

## Usage

Use the [IFM K2 Horizon llama.cpp fork](https://github.com/MBZUAI-IFM/llama.cpp/tree/model/K2Horizon). The example below uses a practical 128K context; use `-c 524288` when available memory permits.

```bash
llama-cli \
  -m K2-Horizon-MoVA-36B-A4B-Q4_K_M.gguf \
  -c 131072 --jinja \
  --temp 1.0 --top-p 0.95
```

## Source

- Source model: [IFM/K2-Horizon-MoVA-36B-A4B](https://huggingface.co/IFM/K2-Horizon-MoVA-36B-A4B)
- Source revision: [`05cab0a`](https://huggingface.co/IFM/K2-Horizon-MoVA-36B-A4B/commit/05cab0a4d7150c1c460a000b37ff40cc1af2feaa)
- Source license: [Apache-2.0](https://www.apache.org/licenses/LICENSE-2.0)