---

base_model: badtheorylabs/BTL-4

tags: [gguf, rocmfp4, llama.cpp, strix-halo, gfx1151, moe, vision, amd, rocm]

pipeline_tag: image-text-to-text

---

BTL-4 — Q4_0_ROCMFP4_STRIX_LEAN (GGUF)

ROCmFP4 quantization of badtheorylabs/BTL-4,

built for AMD Strix Halo (gfx1151) with the

ROCmFPX fork of llama.cpp.

BTL-4 is a 70.2 GB bf16 MoE vision model — Qwen3_5MoeForConditionalGeneration,

40 layers, 256 experts / 8 active, hidden 2048, vocab 248320. An mmproj is included

and is required for image input.

Measured results (Ryzen AI MAX+ 395, 124 GB, gfx1151)

| build | size | real BPW | decode | quality /10 |

|---|---:|---:|---:|---:|

| Q4_0_ROCMFP4_STRIX_LEAN | 18,597,336,928 B (17.32 GiB) | 4.29 | +12% | 9 |

| Q4_K_M (control, run twice) | 21,166,757,728 B (19.71 GiB) | 4.88 | baseline | 8 / 8 |

LEAN is −12% on disk and +12% faster than Q4_K_M, at equal-or-better quality.

Quality used greedy decoding with the baseline run twice to establish the noise floor

(Δ=0 between passes).

A sister build, Q4_0_ROCMFP4_STRIX

(4.31 BPW, 17.39 GiB), was produced independently on a second box and measured +11.4% over

Q4_K_M at 10/10 = 10/10 quality. The two tiers are within ~67 MB and ~0.02 BPW of each other;

prior art on this architecture family favours plain STRIX for quality-critical work.

⚠️ Q3 is a trap on this architecture

| tier | advertised BPW | real BPW |

|---|---:|---:|

| STRIX_LEAN (106) | ~4.38 | 4.29 |

| Q3_0_ROCMFPX | 3.50 | 4.45 |

| Q4_K_M | ~4.58 | 4.88 |

Q3 quantizes larger than LEAN. Always run --dry-run and read the real BPW —

the advertised value is not reliable.

⛔ No speculative decoding

Upstream BTL-4 ships mtp_num_hidden_layers: 0 — there are no MTP tensors.

Do not pass a speculative/MTP flag. A spec flag with no draft tensors loads and reports

READY while producing garbage.

Usage

export HSA_OVERRIDE_GFX_VERSION=11.5.1
export GGML_HIP_ENABLE_UNIFIED_MEMORY=1

llama-server \
  -m BTL-4-Q4_0_ROCMFP4_STRIX_LEAN.gguf \
  --mmproj mmproj-BTL-4-f16.gguf \
  -ngl 999 -fa on -dio --no-warmup --jinja \
  -c 32768 --parallel 1

Requires a ROCmFP4-capable llama.cpp build; stock llama.cpp will not load these tensor types.

License

Inherits the licence of the base model, badtheorylabs/BTL-4.