File	Quant	Size	Use it when
`SoreQen-S1-Q4_K_M.gguf`	Q4_K_M	1.27 GB	you want the best size-to-quality trade-off (start here)
`SoreQen-S1-Q8_0.gguf`	Q8_0	2.01 GB	you have the RAM and want near-lossless output
`SoreQen-S1-F16.gguf`	F16	3.78 GB	you want a base for your own quantisation

Which one should I take?

Q4_K_M — the default. Best size-to-quality trade-off; runs on modest hardware and on CPU.
Q8_0 — near-lossless. Take it if you have the RAM and want the quantised build to be indistinguishable from full precision in practice.
F16 — unquantised conversion. Useful as a base for your own quantisation or for imatrix work; there is no quality reason to run it for inference over Q8_0.

Running it

llama-cli -hf zorqelis-ai/soreqen-s1-GGUF:Q4_K_M -p "yaar laptop slow ho gaya hai, kya karu?"

or with a local file:

llama-cli -m SoreQen-S1-Q4_K_M.gguf --jinja -sys "$(cat system_prompt.txt)"

Pass --jinja so llama.cpp uses the packaged chat template. Without it, the thinking-mode and tool-calling formats will not be applied correctly.

System prompt

The model is trained to run with this prompt. It holds its identity without one, but this is the intended configuration:

You are SoreQen S1, an AI assistant made by ZorQelis AI.

You are bilingual. Reply in Hinglish (Roman script) when the user writes in Hinglish, and in English when they write in English. Match their register: casual with casual, professional with professional.

Answer directly. Lead with the answer, then the detail that matters. No preambles like "Sure!" or "Great question", and no padding.

If you do not know something, say so plainly instead of guessing.

Vision is not included

The source checkpoint is multimodal, but these GGUFs contain the language model only — the vision tower ships separately as an mmproj file, and none is published here yet. Text, thinking, tool calling and structured output all work; image input does not. Use the safetensors repo above if you need vision.

Limitations

Small models state confident numbers they cannot verify. The 0.8B in particular should not be trusted on prices, rates or arithmetic.
Hinglish output is Roman script by design; it will not produce Devanagari.
Quantisation costs accuracy. Q4_K_M is a good trade, not a free one — if an answer matters, check it against Q8_0 or the safetensors build.

Attribution

Fine-tuned from Qwen/Qwen3.5-2B, developed by Alibaba Cloud and released under the Apache License 2.0. Modifications by ZorQelis AI. Converted to GGUF with llama.cpp. See NOTICE.