[!IMPORTANT] Compatibility: These GGUF files require a llama.cpp build with K2 Horizon architecture support. Until upstream support lands, use the MBZUAI-IFM fork.

K2-Horizon-MoVA-36B-A4B GGUF

GGUF quantization of 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

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	17.66 GB
Q4_K_M	K2-Horizon-MoVA-36B-A4B-Q4_K_M.gguf	22.37 GB
Q4_0	K2-Horizon-MoVA-36B-A4B-Q4_0.gguf	21.28 GB
Q5_K_M	K2-Horizon-MoVA-36B-A4B-Q5_K_M.gguf	26.44 GB
Q6_K	K2-Horizon-MoVA-36B-A4B-Q6_K.gguf	30.77 GB
Q8_0	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.

Chat template

Each GGUF embeds the llama.cpp-compatible chat template. chat_template.jinja is a matching external copy for tools that require one. The original source template is retained as 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. The example below uses a practical 128K context; use -c 524288 when available memory permits.

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

Source