K2 Horizon 3.7B needs about 4.9 GB of memory at Q4_K_M (a 3.2 GB file) with an 8K context window, so it fits on a 6 GB GPU. This repository has 6 quants, from 3.2 GB (Q4_K_M) to 10.1 GB (BF16). Longer context needs more: at 32K tokens the KV cache alone adds about 4.8 GB.
K2 Horizon 3.7B – a 5.1B‑parameter dense decoder with 512K context
K2 Horizon 3.7B is a 5.1B‑parameter dense decoder‑only model with a native 524,288‑token context window. It is part of the Horizon family and is fully open, with training data, code, and evaluation resources public. The model has been post‑trained with an anti‑doom‑loop stage to reduce runaway generations.
Good at
- Strong baseline on agentic, coding, and reasoning tasks
- Large 512K context window
- Open‑source training and evaluation
Limits
- No vision support
- Requires llama.cpp with K2 Horizon support
- Large memory for high precision variants
Best for
- Long‑form coding assistance
- Complex reasoning with extended context
- Agentic terminal interactions
Who it’s for. Developers and researchers who need a large‑context, open‑source language model for coding, reasoning, or agentic tasks. The model is suitable for users with access to llama.cpp and sufficient GPU memory.
Quants & memory
| Quant | File | Est. memory | vs VRAM | Result | |
|---|---|---|---|---|---|
| Q4_K_M | 3.2 GB | 4.9 GB | Fits | Get file → | |
| Q5_0 | 3.6 GB | 5.3 GB | Fits | Get file → | |
| Q5_K_M | 3.6 GB | 5.4 GB | Fits | Get file → | |
| Q6_K | 4.2 GB | 5.9 GB | Fits | Get file → | |
| Q8_0 | 5.4 GB | 7.1 GB | Fits | Get file → | |
| BF16 | 10.1 GB | 11.8 GB | Too big | Get file → |
For 8 GB VRAM, use Q4_K_M (3.16 GB file, ~4.9 GB RAM at 8 k context). For 12 GB VRAM, Q5_0 (3.57 GB file, ~5.3 GB RAM) or Q5_K_M (3.64 GB file, ~5.4 GB RAM) are good choices. For 16 GB VRAM, Q6_K (4.16 GB file, ~5.9 GB RAM) or Q8_0 (5.39 GB file, ~7.1 GB RAM) provide higher precision.
What GPU do you need for K2 Horizon 3.7B?
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: the best quant that fits is Q6_K (4.2 GB file, about 5.9 GB in use).
- 8 GB VRAM: the best quant that fits is Q8_0 (5.4 GB file, about 7.1 GB in use).
- 12 GB VRAM: the best quant that fits is BF16 (10.1 GB file, about 11.8 GB in use).
Which K2 Horizon 3.7B quant should you choose?
Q4_K_M is the usual starting point for K2 Horizon 3.7B: it balances file size and quality well. Here it is a 3.2 GB file.
The Q8_0 file is 5.4 GB, about 1.7 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.
New to quant names? Read GGUF quantization explained.
Specifications
Architecture from the GGUF header
Source
Frequently asked questions about K2 Horizon 3.7B
How much VRAM does K2 Horizon 3.7B need?
About 4.9 GB at Q4_K_M with an 8K context window. That is the 3.2 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 3.7B on an 8 GB GPU?
Yes, with Q8_0: the 5.4 GB file needs about 7.1 GB at 8K context. Larger quants in this repository do not fit in 8 GB.
What is the file size of K2 Horizon 3.7B GGUF?
It depends on the quant: Q4_K_M is 3.2 GB, Q5_0 is 3.6 GB, Q5_K_M is 3.6 GB, Q6_K is 4.2 GB, Q8_0 is 5.4 GB, BF16 is 10.1 GB.
What context length does K2 Horizon 3.7B support?
The GGUF header lists a trained context length of 524,288 tokens (512K). The KV cache grows with the context you use: about 1.2 GB at 8K and 4.8 GB at 32K.
Where can I download K2 Horizon 3.7B GGUF?
The files are on Hugging Face in the IFM/K2-Horizon-3.7B-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 3.7B 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 K2 Horizon 3.7B on 8 GB VRAM?
Yes, using the Q4_K_M quantization which requires about 4.9 GB of VRAM at 8 k context.
Does this model support image inputs?
No, K2 Horizon 3.7B is a text‑only decoder model.
What version of llama.cpp is needed?
A version with K2 Horizon architecture support, released in b11454 or later.