
Format	EXL3 safetensors (2 shards, 15.73 GB)
Bits	4.0 bpw body, 6-bit output head, 4-bit MTP
Measured bitrate	4.02 bpw / 6.00 bpw (head)
Converter	ExLlamaV3 1.4.3
Command	`convert.py -b 4.0 -hb 6 -mb 4`
Calibration	250 rows × 2048 cols, `out_scales: always`, codebook `mul1`
Native context	262,144 tokens
Architecture	`Qwen3_5ForConditionalGeneration` — 64 layers, hidden 5120, hybrid Gated DeltaNet (48 linear + 16 full attention)
Load with	ExLlamaV3 ≥ 1.4.3 or TabbyAPI (official ExLlama V3 server)

Not a GGUF, not FP8, not for transformers generate() / vLLM / llama.cpp. The Hub “model size” widget under-counts (packed EXL3 tensors look like ~8B); this is the full 27B.

Sibling quants of the same uncensored source: FP8 · GGUF · MLX. Official (censored) EXL3 of the base: turboderp/Qwen3.8-27B-exl3.

Disclaimer

The parent weights had safety alignment substantially removed by abliteration (refusal-direction orthogonalization). This quant inherits that:

It will comply with requests the original Qwen3.8-27B would refuse.
You are responsible for how you use it and for what it generates.
Apache 2.0 from Qwen/Qwen3.8-27B still applies. Abliteration and this quant do not change the license.
Outputs are not the views of Qwen / Alibaba, OrcaRouter, or the quantizer.

By downloading you accept the above.

What's in the files

Language-model tensors quantized to EXL3 4.0 bpw
Output head at 6 bits (-hb 6)
MTP head quantized at 4 bits (-mb 4) — TabbyAPI draft_mode: mtp works
Vision tower + image/video preprocessor configs (enable vision in the server if you want it; it costs VRAM)
Tokenizer, chat template, generation config

Hardware (measured)

Converted and served on a Razer Blade 16 (2025) RTX 5090 Laptop, 24 GB GDDR7, ExLlamaV3 1.4.3, TabbyAPI, AC power.

VRAM

TabbyAPI settings	GPU memory
Weights only (approx.)	~15.7 GB
`max_seq_len` / `cache_size` 49152, `cache_mode: "8,8"`, `chunk_size: 512`, `max_batch_size: 1`	~16.2 GB / 24 GB
`max_seq_len` / `cache_size` 262144, `cache_mode: Q4`, `chunk_size: 4096`	~21.4 GB / 24 GB

cache_mode is ExLlamaV3 k_bits,v_bits. "8,8" first; "6,6" then "4,4" (legacy Q4) only if you need a longer window on 24 GB. cache_size must be ≥ max_seq_len and a multiple of 256. max_seq_len is prompt + response.

Speed (AC, 16k eval, pre-Q4-cache)

eval/perf.py -cs 16384 -max_length 16384, chunk 4096:

Prefill length	tok/s
256	844
1024	1307
4096	1450
16384	1351

Decode context	tok/s
0	42.1
4096	41.3
8192	40.5
16128	39.2

Use with TabbyAPI

hf download writetoasik/Qwen3.8-27B-Uncensored-exl3-4.0bpw --local-dir ./Qwen3.8-27B-Uncensored-exl3-4.0bpw

In config.yml (24 GB starting point):

yaml

model:
  model_dir: /path/to/models
  model_name: Qwen3.8-27B-Uncensored-exl3-4.0bpw
  backend: exllamav3
  max_seq_len: 49152
  cache_size: 49152
  cache_mode: "8,8"
  chunk_size: 512
  max_batch_size: 1
  vision: false
  reasoning: true
  reasoning_start_token: "<think>"
  reasoning_end_token: "</think>"
  tool_format: qwen3_5
  template_vars_default:
    enable_thinking: true

draft_model:
  draft_mode: mtp

Then:

http://127.0.0.1:5000/v1/chat/completions

If a long prompt appears stuck, drop chunk_size to 256 and keep max_batch_size: 1. Do not set RoPE scale unless you know you need it; this model already trains to 262K.

Need the full native window on 24 GB: max_seq_len / cache_size 262144 and cache_mode: "4,4" (or Q4). Quality of the KV cache is worse than "8,8".

Use with ExLlamaV3 directly

python

from exllamav3 import Config, Model, Tokenizer, Generator, Job

model_dir = "Qwen3.8-27B-Uncensored-exl3-4.0bpw"
cfg = Config.from_directory(model_dir)
model = Model.from_config(cfg)
model.load()
tok = Tokenizer.from_config(cfg)
gen = Generator(model, tok)

prompt = tok.encode("Hello.")
job = Job(input_ids=prompt, max_new_tokens=128)
gen.enqueue(job)
print(tok.decode(gen.iterate()[0]["token_ids"]))

See ExLlamaV3 examples.

Conversion

text

python convert.py \
  -i orcarouter/Qwen3.8-27B-Uncensored \
  -w ./_exl3_work \
  -o ./Qwen3.8-27B-Uncensored-exl3-4.0bpw \
  -b 4.0 -hb 6 -mb 4

Flag	Value	Meaning
`-b`	4.0	body bits per weight
`-hb`	6	lm_head bits
`-mb`	4	MTP bits

quantization_config.json in this repo is the converter's record (method exl3, version 1.4.3).

Parent model (abliteration, eval)

All refusal / capability numbers below are OrcaRouter's measurements on the BF16 (and its FP8), not re-run on this EXL3. Treat them as properties of the source, not a claim that 4.0 bpw is bit-identical in quality.

Refusal, thinking off (lower = less refusal): AdvBench 0.0%, JailbreakBench 0.0%, StrongREJECT 2.0%, HarmBench 2.7%, MaliciousInstruct 0.0% vs 94–99% on stock Qwen3.8-27B.

Capability vs stock (same scripts): MMLU 84.7% (+0.4), MMLU-Pro 76.8% (−0.8), GSM8K 88.7% (−1.3), CMMLU 80.8% (−0.6). WikiText-2-raw PPL 6.96 on the BF16.

Full write-up: orcarouter/Qwen3.8-27B-Uncensored.

License

Apache 2.0, inherited from Qwen/Qwen3.8-27B.

Base: Qwen / Alibaba Cloud
Abliteration: OrcaRouter
EXL3 quant: writetoasik

Redistribution must keep the Apache 2.0 license text (this repo includes LICENSE).