Qwen3.8-27B-EXL3-3.0bpw

Qwen3.8-27B — EXL3 3.0bpw (mul1)

This is an EXL3 quantization of Qwen/Qwen3.8-27B, a 27B dense hybrid-GDN vision-language model. EXL3 uses QTIP-style trellis quantization, which compresses far more aggressively than round-to-nearest GPTQ/AWQ-style methods while holding up better at low bitrates.

• Codebook: mul1
• Bits per weight: 3.0 (linear layers), 6 bits (lm_head and vision tower), 4 bits (MTP head)
• Calibration: 256 rows × 2048 cols
• On-disk size: ~13 GB (down from ~54 GB BF16)
• Base architecture: Qwen3_5MoeForConditionalGeneration (hybrid linear/full attention, native vision-language)

Quality validation

Quantization quality was validated against the original BF16 weights via KL-divergence and perplexity on a held-out evaluation set, not just calibration-set loss:

The quantized model's perplexity essentially matches the original — the low KLD and high top-1 agreement indicate the 3.0bpw trellis quantization preserves the model's output distribution closely, not just its loss on average.

Usage

Requires exllamav3 (or a frontend built on it, e.g. TabbyAPI):

pip install exllamav3

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

config = Config.from_directory("path/to/this/repo")
model = Model.from_config(config)
cache = Cache(model, max_num_tokens=8192)
model.load()
tokenizer = Tokenizer.from_config(config)

generator = Generator(model=model, cache=cache, tokenizer=tokenizer)
output = generator.generate(prompt="Explain the significance of the Riemann hypothesis.", max_new_tokens=512)
print(output)

Original model card

For architecture details, benchmark numbers, and usage guidance for the base model, see Qwen/Qwen3.8-27B.

License

Apache 2.0, inherited from the base model.