Qwen3.6-27B-MTP-ROCmFP4_FAST-GGUF

Qwen3.6-27B-MTP ROCmFP4_FAST — GGUF

ROCmFP4_FAST quant of Qwen/Qwen3.6-27B (Apache 2.0), produced via charlie12345/ROCmFPX's ROCm fork. Benchmarked on an RDNA4 RX 9060 XT (16 GB). Mean AEON score: 0.558 (120 scored / 150 cases). Companion TQ3_4S quant for Blackwell CUDA is in a separate repo.

File

NOT a stock llama.cpp quant

ROCmFP4_FAST is a custom weight format unique to charlie12345/ROCmFPX. Stock llama.cpp and nixpkgs llama-cpp will exit with unknown quantization at load time. Use the llama-server/llama-cli from the ROCmFPX fork.

Scope of these benchmarks — read this first

These numbers are a light baseline, not a thorough ROCmFPX evaluation. The mesh's bench framework is built for production agent workload regression-detection on the local stack, not for the kind of multi-axis sweep that upstream quant maintainers typically publish. Specifically:

• Harness scope is bounded. The numbers below come from the mesh's Aeon-Bench-Pod (self-reported mode, 150-case full suite, text-only, no agentic harness). That's a regression suite, not a quality benchmark.
• Sample sizes are small. 150 cases on a single GPU, single rep. None are powered for statistical significance.
• No perplexity / wikitext / MMLU / GSM8K. The mesh's stack isn't a quality benchmark — those are upstream's territory.
• Single GPU class (RDNA4 16 GB). All measurements are on an AMD RX 9060 XT 16 GB (gfx1200, ROCm 7.x, charlie12345/ROCmFPX). No Blackwell, no CDNA, no multi-GPU. Cross-hardware generalization is NOT implied.
• No human eval. "0.558 mean on the AEON suite" is not a quality verdict on this specific quant.
• 30 prose cases unscored — no frontier judge endpoint configured, so the prose category (30/150 cases) is excluded from the mean.

What this IS good for: a quick signal that the quant (a) loads, (b) runs at production-usable throughput, (c) produces coherent output across the main categories. What this is NOT good for: claiming "this is the best quant of this model," reproducing academic benchmark results, or substituting for upstream's validation work.

For a rigorous view, see Qwen/Qwen3.6-27B (parent model) and charlie12345/ROCmFPX (quantizer). Raw bench reports are attached as BENCH-*.md files in this repo.

What we measured

AEON v3 — 150 cases (120 scored)

Benchmarked on AMD RDNA4 RX 9060 XT 16 GB, full suite.

Profile: Strong coding (0.900), mid-range math (0.533) and reasoning (0.467), weak on instruction-following (0.333). The ROCmFP4_FAST arm scores slightly higher than the companion TQ3_4S (0.525) due to stronger math and reasoning performance on this hardware.

Companion: TQ3_4S (Blackwell)

The companion TQ3_4S quant for Blackwell CUDA scores 0.525 on the same suite. Both formats produce similar profile shapes; the AMD arm scores ~3pp higher on mean.

Quick start

# Build charlie12345/ROCmFPX (ROCm fork)
git clone https://github.com/charlie12345/ROCmFPX
cd ROCmFPX
mkdir build && cd build
cmake .. -DGGML_HIP=ON -DAMDGPU_TARGETS=gfx1200
make -j$(nproc)

# Serve
llama-server \
  -m Qwen3.6-27B-MTP-ROCmFP4_FAST.gguf \
  --host 0.0.0.0 --port 8081 \
  -ngl 99 -c 65536 -t 12 \
  -ctk q4_0 -ctv q4_0 \
  -fa on --cache-ram 0 --no-cache-prompt \
  -np 1 --batch-size 512 --ubatch-size 128 \
  --jinja --metrics -rea off

Reproduce the quant

# Requires the ROCmFPX fork and the BF16 source GGUF
llama-quantize --allow-requantize Qwen3.6-27B-BF16.gguf \
  Qwen3.6-27B-MTP-ROCmFP4_FAST.gguf Q4_0_ROCMFP4_FAST

Files in this repo

What's NOT in this repo (caveats)

• Stock llama.cpp will not load this file. ROCmFP4_FAST is unique to charlie12345/ROCmFPX.
• No CUDA / non-AMD GPU bench. All measurements are RDNA4 (gfx1200). The companion TQ3_4S quant for Blackwell is in a separate repo.
• No quality benchmark (perplexity, MMLU, GSM8K). The custom 4-bit quant works on the mesh's regression tests; whether it's "the best ROCmFPX quant" needs upstream validation.
• No MTP / speculative-decode bench. MTP heads are present in the source model but were not benched on this quant.

Provenance

• Source model: Qwen/Qwen3.6-27B — Apache 2.0
• Quantizer: charlie12345/ROCmFPX
• Quantizer license: MIT
• Build hardware: AMD RX 9060 XT 16 GB (RDNA4, gfx1200), ROCm 7.x, NixOS 25.11
• Bench harness: Aeon-Bench-Pod v1 (self-reported mode)

License

• Qwen/Qwen3.6-27B is Apache 2.0.
• charlie12345/ROCmFPX is MIT.
• The GGUF in this repo is a derivative of the Apache 2.0-licensed parent, produced with the MIT-licensed quantizer. The Apache 2.0 license is preserved.