[!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-7B GGUF

GGUF quantizations of IFM/K2-Horizon-7B, a 7B dense decoder-only model for reasoning, coding, long-context work, and tool use. The source checkpoint supports a native context length of 524,288 tokens (512K).

Benchmarks

Benchmark results reported by IFM for the original K2-Horizon-7B checkpoint.

GGUF files

Quantization	File	Size
Q4_0	K2-Horizon-7B-Q4_0.gguf	5.34 GB
Q4_K_M	K2-Horizon-7B-Q4_K_M.gguf	5.59 GB
Q4_K_M Selective	K2-Horizon-7B-Q4_K_M-Selective.gguf	5.96 GB
Q5_K_M	K2-Horizon-7B-Q5_K_M.gguf	6.47 GB
Q6_K	K2-Horizon-7B-Q6_K.gguf	7.39 GB
Q8_0	K2-Horizon-7B-Q8_0.gguf	9.57 GB

The files are text-only GGUFs; no vision projector is required. The selective variant uses a Q4_K_M baseline with attention Q/K/V/O projection tensors kept at Q6_K; it is a manual tensor-selective build and does not use an importance matrix. 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; -c 524288 can be used when the available memory is sufficient.

llama-cli \
  -m K2-Horizon-7B-Q4_K_M.gguf \
  -c 131072 --jinja \
  --temp 1.0 --top-p 0.95

Source