K2-Horizon-3.7B — GGUF Quantizations

GGUF quantizations of IFM/K2-Horizon-3.7B, a dense 3.7B-parameter causal decoder (K2HorizonForCausalLM, model_type: k2_horizon).

Quantized by NANI-Nithin using a custom pipeline built on the MBZUAI-IFM llama.cpp fork (branch model/K2Horizon).

Note on file naming: IFM's official GGUF repo ships this model as K2-Horizon-4B-BF16.gguf (rounded name). Files in this repo are renamed to the K2-Horizon-3.7B-* convention for consistency.

Model Details

Property	Value
Base model	IFM/K2-Horizon-3.7B
Architecture	`K2HorizonForCausalLM` (`k2_horizon`)
Parameters	~3.7B (dense decoder, no MoE)
Original dtype	BF16
BF16 GGUF size	~10.13 GB
Source GGUF	IFM/K2-Horizon-3.7B-GGUF
llama.cpp fork	MBZUAI-IFM/llama.cpp @ model/K2Horizon

Note: These GGUFs carry the k2-horizon architecture token and require the MBZUAI-IFM fork (or upstream llama.cpp once support is merged) to run. Vanilla upstream llama.cpp (as of September 2026) does not support K2HorizonForCausalLM.

Included Files

Standard Quantizations

File	Bits/Weight	Notes
`K2-Horizon-3.7B-BF16.gguf`	16 bpw	Source quant, full precision
`K2-Horizon-3.7B-Q8_0.gguf`	8 bpw	Near-lossless, recommended reference
`K2-Horizon-3.7B-Q6_K.gguf`	6 bpw	Near-lossless K-quant
`K2-Horizon-3.7B-Q5_K_M.gguf`	5 bpw	Best quality/size in the 5-bit range
`K2-Horizon-3.7B-Q5_K_S.gguf`	5 bpw	Smaller 5-bit variant
`K2-Horizon-3.7B-Q5_1.gguf`	5 bpw	Legacy 5-bit
`K2-Horizon-3.7B-Q5_0.gguf`	5 bpw	Legacy 5-bit
`K2-Horizon-3.7B-Q4_K_M.gguf`	4 bpw	Recommended general use
`K2-Horizon-3.7B-Q4_K_S.gguf`	4 bpw	Smaller 4-bit K-quant
`K2-Horizon-3.7B-Q4_1.gguf`	4 bpw	Legacy 4-bit
`K2-Horizon-3.7B-Q4_0.gguf`	4 bpw	Legacy 4-bit
`K2-Horizon-3.7B-Q3_K_L.gguf`	3 bpw	Large 3-bit K-quant
`K2-Horizon-3.7B-Q3_K_M.gguf`	3 bpw	Medium 3-bit K-quant
`K2-Horizon-3.7B-Q3_K_S.gguf`	3 bpw	Small 3-bit K-quant
`K2-Horizon-3.7B-Q2_K.gguf`	2 bpw	Aggressive compression
`K2-Horizon-3.7B-Q2_K_S.gguf`	2 bpw	Smaller 2-bit K-quant (imatrix-guided)
`K2-Horizon-3.7B-Q2_0.gguf`	2.25 bpw	Group-64 2-bit
`K2-Horizon-3.7B-Q1_0.gguf`	1.125 bpw	Maximum compression

IQ (Importance-Matrix) Quantizations

File	Bits/Weight
`K2-Horizon-3.7B-IQ4_NL.gguf`	~4 bpw
`K2-Horizon-3.7B-IQ4_XS.gguf`	~4 bpw
`K2-Horizon-3.7B-IQ3_M.gguf`	~3 bpw
`K2-Horizon-3.7B-IQ3_S.gguf`	~3 bpw
`K2-Horizon-3.7B-IQ3_XS.gguf`	~3 bpw
`K2-Horizon-3.7B-IQ3_XXS.gguf`	~3 bpw
`K2-Horizon-3.7B-IQ2_M.gguf`	~2 bpw
`K2-Horizon-3.7B-IQ2_S.gguf`	~2 bpw
`K2-Horizon-3.7B-IQ2_XS.gguf`	~2 bpw
`K2-Horizon-3.7B-IQ2_XXS.gguf`	~2 bpw
`K2-Horizon-3.7B-IQ1_M.gguf`	1.75 bpw
`K2-Horizon-3.7B-IQ1_S.gguf`	1.56 bpw

Quantization Method

Source: IFM's official BF16 GGUF (originally named K2-Horizon-4B-BF16.gguf, renamed here).
imatrix: Computed from Salesforce/wikitext (wikitext-2-raw-v1, 500 rows) with 24 GPU layers offloaded on an RTX 4060 Laptop (8 GB VRAM). Applied to all K-quants below Q6 and all IQ quants.
Fork: MBZUAI-IFM/llama.cpp, branch model/K2Horizon.

Usage

Requires the MBZUAI-IFM llama.cpp fork (model/K2Horizon branch).

git clone -b model/K2Horizon https://github.com/MBZUAI-IFM/llama.cpp
cd llama.cpp && cmake -B build -DGGML_CUDA=ON && cmake --build build --config Release

./build/bin/llama-cli \
  -m K2-Horizon-3.7B-Q4_K_M.gguf \
  -p "Hello, I am" \
  -n 128

License

Weights are released under the same license as the original IFM/K2-Horizon-3.7B model. Please refer to the original repository for full license terms.

Credits

Original model: MBZUAI IFM
Quantized by: NANI-Nithin
Quantization tool: MBZUAI-IFM/llama.cpp

NANI-Nithin/K2-Horizon-3.7B-GGUF