Most quantized weights ship with an adjective. Ours ship with a table.

Head to head with unsloth/Qwen3.5-9B-GGUF, same Q4_K_M rung — KL divergence from bf16, lower is better:
Korean −25.4 % · English −11.4 %
_{Two axes measured, two ahead. What we did not win is further down, with its number.}

📚 Collection

▶ Archsloth MODELs — every build in the ARCHsloth Q4 grade Qwen3-4B · Qwen3.5-9B · Qwen3.8-27B

Qwen3.5-9B · ARCHsloth Q4

Same rung. Same tensor types. Only the rounding differs.

Which file do I want?

File	Size	What it is
`Qwen3.5-9B-AS-Q4_K_M.gguf` ⭐	5.78 GB	The weights. This is the model.
`mmproj-F32.gguf`	1.82 GB	The vision tower. Only needed if you want the model to see images.

Qwen3.5-9B is a vision-language model. Text-only chat needs the first file alone; pass --mmproj mmproj-F32.gguf to llama-server to turn the eyes on. The vision tower is not quantized — it is carried at full precision, exactly as the base model shipped it.

`[measured]` Q4_K_M — the same rung, both sides

KL divergence from the bf16 original. Lower is better. llama-perplexity --kl-divergence, ctx 512, 60 chunks, held-out Korean and English text that is not in the calibration set.

	bytes	Korean	English
ARCHsloth `Q4_K_M`	5,780,090,240	0.019255	0.019237
unsloth `Q4_K_M`	5,680,522,464	0.025821	0.021718
unsloth `UD-Q4_K_XL`	5,966,095,584	0.013740	0.012546
stock `llama-quantize`	5,780,090,304	0.043475	0.037879

Against	Korean	English
unsloth `Q4_K_M` — same rung	−25.4 %	−11.4 %
stock `llama-quantize`	−55.7 %	−49.2 %

Same-top-p moves the same direction on every row (93.974 % vs 92.784 % on Korean). Full statistics: EVAL.md. Raw per-run logs: eval/logs/.

Quickstart

# text only
llama-cli -m Qwen3.5-9B-AS-Q4_K_M.gguf -p "안녕하세요" -ngl 0 -t 8

# OpenAI-compatible server, with vision
llama-server -m Qwen3.5-9B-AS-Q4_K_M.gguf --mmproj mmproj-F32.gguf \
  --host 127.0.0.1 --port 8080 --jinja

No fork. No custom format. The layout is llama.cpp's own, which is the point.

What is actually different

--scheme GGUF:Q4_K_M     optimize the rounding for the quantizer that is used at export
--enable_alg_ext         sign-gradient rounding search (SignRoundV2, arXiv:2512.04746)
--dataset cal_archsloth.jsonl
--iters 200  --nsamples 128

Nothing else. Bit widths and tensor types are llama.cpp's standard Q4_K_M. Nothing below 4 bits anywhere.

The calibration set — it ships in this repo

cal_archsloth.jsonl, 200 samples of exactly 9,000 characters each, Korean and English interleaved at the sample level. Interleaving is what mattered, not the ratio.

This is the lever. The same model built with an English-only corpus scored 0.027336 on Korean — worse than the file we are comparing against. Same code, same flags, same bit widths; only the calibration text changed, and it moved Korean by 29.6 %. That build is not in this repository. Details: CALIBRATION.md.

What we did not win

[measured] unsloth's UD-Q4_K_XL is ahead of this file. 0.013740 Korean and 0.012546 English against our 0.019255 / 0.019237 — they are 40.1 % and 53.3 % better. It is a 3.2 % larger file with a different tensor map, and on this model that map is worth more than our rounding search. We publish the row because a number you omit is a number someone else produces. Closing it is the open work on this model.
[measured] Two axes, not ten. Korean and English are measured here. The ten-axis sweep that the 4B card carries has not been run on this model. We publish only what we ran ourselves.
[pending] Q6_K and Q8_0 rungs, long-context retrieval, agentic tool use, and the vision axis. Not measured on this model.

Reproduce this

# 1. the comparison file, straight from its own repository
hf download unsloth/Qwen3.5-9B-GGUF Qwen3.5-9B-Q4_K_M.gguf --local-dir rival

# 2. the reference the distance is measured from
python llama.cpp/convert_hf_to_gguf.py Qwen/Qwen3.5-9B --outtype bf16 --outfile ref-BF16.gguf

# 3. teacher logits, once per evaluation text
llama-perplexity -m ref-BF16.gguf -f eval/eval_ko.txt \
  --kl-divergence-base base_ko.dat -ngl 99 -c 512 --chunks 60

# 4. every candidate against the same teacher, same text, same chunks
llama-perplexity -m <candidate>.gguf -f eval/eval_ko.txt \
  --kl-divergence --kl-divergence-base base_ko.dat -ngl 99 -c 512 --chunks 60

If a number here does not replicate on your hardware, open a discussion and we will correct it.

Lineage

Component	Origin
Weights	Qwen/Qwen3.5-9B, unmodified — quantized from the original, not from anyone else's quantization
Layout, bit widths, tensor types	llama.cpp standard `Q4_K_M`, unchanged
Vision tower	carried at full precision, not quantized
Runtime	upstream llama.cpp, unmodified
Rounding search	Intel AutoRound with `enable_alg_ext` (SignRoundV2)
Calibration corpus, evaluation set, and the measurement	ours — all three ship in this repo

License

Apache-2.0, inherited from the base model. Comparison files are named so you can fetch the exact bytes we measured.

Archsloth/Qwen3.5-9B-GGUF

📚 Collection