Archsloth/Qwen3.8-27B-GGUF

🤗 Hugging Face sourceimage-text-to-textapache-2.024 GBGGUFHF checksums availableupdated today
No torrent yet

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

Head to head with unsloth/Qwen3.8-27B-GGUF, same Q6_K rung — KL divergence from bf16, lower is better:
Korean −36.9 % · English −15.0 %
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

Qwen3.8-27B · ARCHsloth Q6

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


Which file do I want?

File Size What it is
Qwen3.8-27B-AS-Q6_K.gguf 22.43 GB The weights. This is the model.
mmproj-F32.gguf 1.84 GB The vision tower. Only needed if you want the model to see images.

Qwen3.8-27B 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.

There is no Q4 file in this repository. Why not.


[measured] Q6_K — 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 Q6_K 22,430,999,840 0.001385 0.001254
unsloth UD-Q6_K 21,983,677,344 0.002195 0.001476
unsloth UD-Q6_K_M 23,088,409,504 0.001632 0.001137
Against Korean English
unsloth UD-Q6_K — same rung −36.9 % −15.0 %
unsloth UD-Q6_K_M — 2.9 % larger than ours −15.1 % +10.3 %

Same-top-p moves the same direction as the divergence on every row (98.222 % vs 97.693 % on Korean). Full statistics: EVAL.md. Raw per-run logs: eval/logs/.


Quickstart

# text only
llama-cli -m Qwen3.8-27B-AS-Q6_K.gguf -p "안녕하세요" -ngl 99

# OpenAI-compatible server, with vision
llama-server -m Qwen3.8-27B-AS-Q6_K.gguf --mmproj mmproj-F32.gguf \
  --host 127.0.0.1 --port 8080 --jinja

A 27B is a dense model: every token reads the whole file. Plan for the weights plus the KV cache, and the cache is what decides whether a long-context session fits.


What is actually different

--scheme GGUF:Q6_K       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 Q6_K. Nothing below 4 bits anywhere.

The calibration set — it ships in this repo

cal_archsloth.jsonl, 500 samples of exactly 9,000 characters each, Korean and English interleaved at the sample level. Interleaving is what mattered, not the ratio. Details: CALIBRATION.md.


What we did not win

  • [measured] Our Q4_K_M lost English, so it is not in this repository. Against Qwen3.8-27B-UD-Q4_K_M.gguf it took Korean by −12.8 % and gave English away by +15.8 %. A rung we do not win is a rung we do not ship. It returns when it wins.
  • [measured] UD-Q6_K_M keeps English. That file is 2.9 % larger than ours. We take Korean from it by −15.1 % and it takes English back by +10.3 %. Against the file at our own rung, UD-Q6_K, we are ahead on both.
  • [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] Long-context retrieval, agentic tool use, and the vision axis.

Reproduce this

# 1. the comparison file, straight from its own repository
hf download unsloth/Qwen3.8-27B-GGUF Qwen3.8-27B-UD-Q6_K.gguf --local-dir rival

# 2. the reference the distance is measured from
python llama.cpp/convert_hf_to_gguf.py Qwen/Qwen3.8-27B --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/Qwen3.8-27B, unmodified — quantized from the original, not from anyone else's quantization
Layout, bit widths, tensor types llama.cpp standard Q6_K, 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.