HO

K2 Horizon MoVA 36B A4B

HorizonGGUF by NANI-NithinUpdated 2026-09-05
Download on Hugging Face
Parameters37.4B
Context512K
Layers48
Architecturek2-horizon
Smallest quant5.8 GB
Quick answer

K2 Horizon MoVA 36B A4B needs about 24.5 GB of memory at Q4_K_M (a 22.4 GB file) with an 8K context window, so it fits on a 32 GB GPU. This repository has 31 quants, from 5.8 GB (Q1_0) to 74.9 GB (BF16). Longer context needs more: at 32K tokens the KV cache alone adds about 6.4 GB.

K2 Horizon MoVA 36B A4B – a sparse 36B parameter MoE model with MoVA attention

K2 Horizon MoVA 36B A4B is a 36 billion‑parameter Mixture‑of‑Experts model that activates only 4 billion parameters per token. It supports a native 524,288‑token context window and is released under the Apache‑2.0 license.

Good at

  • Advanced reasoning
  • Large context handling
  • Sparse expert routing

Limits

  • Requires MBZUAI‑IFM fork of llama.cpp
  • No vision support
  • Large VRAM needed for higher quants

Best for

  • Complex reasoning tasks
  • Long‑form document analysis
  • Research prototyping

Who it’s for. Developers and researchers needing high‑capacity reasoning with sparse activation. The model is fully open source and suitable for experimentation on high‑end GPUs.

Quants & memory

QuantFileEst. memoryvs VRAMResultDownload
Q1_05.8 GB 7.9 GB
Tight Get file →
IQ1_S7.9 GB 10.0 GB
Too big Get file →
IQ1_M8.7 GB 10.8 GB
Too big Get file →
IQ2_XXS10.1 GB 12.2 GB
Too big Get file →
Q2_K_S11.1 GB 13.2 GB
Too big Get file →
Q2_011.1 GB 13.2 GB
Too big Get file →
IQ2_XS11.2 GB 13.3 GB
Too big Get file →
IQ2_S11.4 GB 13.5 GB
Too big Get file →
IQ2_M12.5 GB 14.6 GB
Too big Get file →
Q2_K13.6 GB 15.7 GB
Too big Get file →
IQ3_XXS14.6 GB 16.7 GB
Too big Get file →
IQ3_XS15.7 GB 17.8 GB
Too big Get file →
Q3_K_S16.4 GB 18.5 GB
Too big Get file →
IQ3_S16.4 GB 18.5 GB
Too big Get file →
IQ3_M16.5 GB 18.7 GB
Too big Get file →
Q3_K_M17.7 GB 19.8 GB
Too big Get file →
Q3_K_L18.8 GB 21.0 GB
Too big Get file →
IQ4_XS20.1 GB 22.2 GB
Too big Get file →
Q4_021.3 GB 23.4 GB
Too big Get file →
IQ4_NL21.3 GB 23.4 GB
Too big Get file →
Q4_K_S21.4 GB 23.5 GB
Too big Get file →
MXFP4_MOE21.7 GB 23.8 GB
Too big Get file →
Q4_K_M22.4 GB 24.5 GB
Too big Get file →
Q4_123.6 GB 25.7 GB
Too big Get file →
Q5_025.9 GB 28.0 GB
Too big Get file →
Q5_K_S25.9 GB 28.0 GB
Too big Get file →
Q5_K_M26.4 GB 28.6 GB
Too big Get file →
Q5_128.2 GB 30.3 GB
Too big Get file →
Q6_K30.8 GB 32.9 GB
Too big Get file →
Q8_039.8 GB 41.9 GB
Too big Get file →
BF1674.9 GB 77.0 GB
Too big Get file →

For 8 GB VRAM, use Q1_0 (5.76 GB) or IQ1_S (7.86 GB). For 16 GB VRAM, Q2_K_S (11.1 GB) or IQ2_S (11.36 GB) are suitable. For 24 GB VRAM, Q4_K_M (22.37 GB) or Q4_1 (23.58 GB) provide a balance of size and performance.

What GPU do you need for K2 Horizon MoVA 36B A4B?

These figures add the file size, an fp16 KV cache at an 8K context window (worked out from this model’s own header) and about 0.5 GB for the runtime. They are estimates, so leave a little headroom, and expect to use less context or a smaller quant if you are right at the limit.

  • 4 GB VRAM: no quant fits at an 8K context.
  • 6 GB VRAM: no quant fits at an 8K context.
  • 8 GB VRAM: the best quant that fits is Q1_0 (5.8 GB file, about 7.9 GB in use).
  • 12 GB VRAM: the best quant that fits is IQ1_M (8.7 GB file, about 10.8 GB in use).
  • 16 GB VRAM: the best quant that fits is Q2_K (13.6 GB file, about 15.7 GB in use).
  • 24 GB VRAM: the best quant that fits is MXFP4_MOE (21.7 GB file, about 23.8 GB in use).
  • 32 GB VRAM: the best quant that fits is Q5_1 (28.2 GB file, about 30.3 GB in use).
  • 48 GB VRAM: the best quant that fits is Q8_0 (39.8 GB file, about 41.9 GB in use).

Which K2 Horizon MoVA 36B A4B quant should you choose?

Q4_K_M is the usual starting point for K2 Horizon MoVA 36B A4B: it balances file size and quality well. Here it is a 22.4 GB file.

The Q8_0 file is 39.8 GB, about 1.8 times the size of Q4_K_M. It keeps the weights closer to the original, which only matters if you have the memory to spare.

If memory is tight, the smallest file is Q1_0 at 5.8 GB, with the quality trade-offs that come with fewer bits. Lower bit counts mean smaller files and more quality loss.

New to quant names? Read GGUF quantization explained.

Specifications

Architecture from the GGUF header

Architecturek2-horizon
Parameters37.4B (approximate, mixture of experts)
Active parameters4B
Layers48
Trained context524,288 tokens
KV cache (fp16)192 KB per token

Source

Quantized byNANI-Nithin
FamilyHorizon
Licenseapache-2.0
Downloads (HF)14,731
Last updated2026-09-05

Frequently asked questions about K2 Horizon MoVA 36B A4B

How much VRAM does K2 Horizon MoVA 36B A4B need?

About 24.5 GB at Q4_K_M with an 8K context window. That is the 22.4 GB file, plus the KV cache and roughly 0.5 GB of runtime overhead. Smaller quants need less and longer context needs more.

Can I run K2 Horizon MoVA 36B A4B on an 8 GB GPU?

Yes, with Q1_0: the 5.8 GB file needs about 7.9 GB at 8K context. Larger quants in this repository do not fit in 8 GB.

What is the file size of K2 Horizon MoVA 36B A4B GGUF?

It depends on the quant: Q1_0 is 5.8 GB, IQ1_S is 7.9 GB, IQ1_M is 8.7 GB, IQ2_XXS is 10.1 GB, Q2_K_S is 11.1 GB, Q2_0 is 11.1 GB and 25 more.

What context length does K2 Horizon MoVA 36B A4B support?

The GGUF header lists a trained context length of 524,288 tokens (512K). The KV cache grows with the context you use: about 1.6 GB at 8K and 6.4 GB at 32K.

Where can I download K2 Horizon MoVA 36B A4B GGUF?

The files are on Hugging Face in the NANI-Nithin/K2-Horizon-MoVA-36B-A4B-GGUF repository. The download buttons on this page open the file on Hugging Face. ggufmodels.com does not host any model files.

What license does K2 Horizon MoVA 36B A4B use?

Hugging Face lists the license as apache-2.0. Licenses can differ between the quantized files and the original model, so check the repository and the base model before commercial use.

Can I run this model on 8 GB VRAM?

Yes, using the Q1_0 or IQ1_S quantizations, which require about 7.9–10 GB of VRAM at 8k context.

Does this model support image input?

No, the model has no vision capabilities and only handles text.

What fork of llama.cpp is required?

The MBZUAI‑IFM fork of llama.cpp with native k2‑horizon support is necessary to load these GGUFs.

Other GGUFs of K2-Horizon-MoVA-36B-A4B

Similar models

Browse all →

More from NANI-Nithin

All by NANI-Nithin →

More Horizon models

All Horizon →

Other large models

All large models →