---

license: apache-2.0

base_model: Qwen/Qwen3.8-27B

tags:

• gguf
• 1bit-engine
• strix-halo
• rocm
• hadamard
• w4a4

pipeline_tag: text-generation

---

Qwen3.8-27B Q4_0-H32 (Hadamard-rotated) for the 1bit engine

Qwen/Qwen3.8-27B as a Q4_0 GGUF whose matmul weights

are rotated by a 32-point Walsh-Hadamard transform. It is made for the prompt-processing route of

the 1bit engine on AMD Strix Halo (Ryzen AI Max+ 395,

Radeon 8060S): 4-bit weights times 4-bit activations on the GPU's matrix units, with the rotation

keeping the 4-bit activations close to the full model.

It only runs correctly on the 1bit engine's lean ROCm build. Its weights are rotated, so the

activations must be rotated the same way, and only that build does it. Stock llama.cpp loads the

file but computes garbage. 1bit serve recognises the file (it carries

onebit.hadamard_q4_0 = 32) and picks the right route by itself.

Run it

# engine built with -DONEBIT_LEAN=ON -DONEBIT_LEAN_ROCM=ON (docs/lean.md)
1bit serve -m Qwen3.8-27B-Q4_0-H32.gguf

Measured (Strix Halo, 2026-09-28)

KLD against BF16 over wikitext-2 (40 x 512); pp512 from llama-bench (ub 512):

| Qwen3.8-27B | pp512 | PPL | Mean KLD | Same top token |

|---|---|---|---|---|

| Q4_0, exact int8 | ~400 | 6.021 | 0.029 | 91.9% |

| Q4_0, W4A4 without rotation | 461-488 | 6.241 | 0.084 | 87.4% |

| this file, W4A4 | 509 | 6.096 | 0.055 | 89.3% |

Through 1bit serve: a 1,838-token prompt at 440-470 tok/s. Decode on this route is about

11 tok/s; for decode-heavy work run Unsloth's Qwen3.8-27B GGUF on the engine's Vulkan route with

the DFlash2 drafter (1bit serve --device vulkan --dflash ..., 45.7 tok/s on code).

How it was made

tools/hadamard_q4_0.py Qwen3.8-27B-Q8_0.gguf Qwen3.8-27B-Q4_0-H32.gguf --imatrix imatrix_unsloth.gguf

From unsloth/Qwen3.8-27B-GGUF's Q8_0 and its

imatrix. Rotated and Q4_0: the attention, FFN and delta-net alpha/beta projections (456 tensors).

ssm_out is Q5_K, the output head Q6_K, the MTP projection Q8_0, norms and convolutions F32.

15,182 MiB (4.66 bits per weight).

sha256 fd61323dd0f78deeb65f7af5a5b2dcd8468af1edb531a9cd37d156bf49e37df4

Licence

Apache-2.0, as the base model (LICENSE). Qwen3.8 by the Qwen team. Quantization

source and imatrix by Unsloth.