Instructions to use maczzzzzz/Qwen3.6-27B-MTP-ROCmFP4_FAST-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use maczzzzzz/Qwen3.6-27B-MTP-ROCmFP4_FAST-GGUF with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="maczzzzzz/Qwen3.6-27B-MTP-ROCmFP4_FAST-GGUF",
	filename="Qwen3.6-27B-MTP-ROCmFP4_FAST.gguf",
)

llm.create_chat_completion(
	messages = "No input example has been defined for this model task."
)

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use maczzzzzz/Qwen3.6-27B-MTP-ROCmFP4_FAST-GGUF with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf maczzzzzz/Qwen3.6-27B-MTP-ROCmFP4_FAST-GGUF
# Run inference directly in the terminal:
llama cli -hf maczzzzzz/Qwen3.6-27B-MTP-ROCmFP4_FAST-GGUF

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf maczzzzzz/Qwen3.6-27B-MTP-ROCmFP4_FAST-GGUF
# Run inference directly in the terminal:
llama cli -hf maczzzzzz/Qwen3.6-27B-MTP-ROCmFP4_FAST-GGUF

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf maczzzzzz/Qwen3.6-27B-MTP-ROCmFP4_FAST-GGUF
# Run inference directly in the terminal:
./llama-cli -hf maczzzzzz/Qwen3.6-27B-MTP-ROCmFP4_FAST-GGUF

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf maczzzzzz/Qwen3.6-27B-MTP-ROCmFP4_FAST-GGUF
# Run inference directly in the terminal:
./build/bin/llama-cli -hf maczzzzzz/Qwen3.6-27B-MTP-ROCmFP4_FAST-GGUF

Use Docker

docker model run hf.co/maczzzzzz/Qwen3.6-27B-MTP-ROCmFP4_FAST-GGUF

LM Studio
Jan
Ollama
How to use maczzzzzz/Qwen3.6-27B-MTP-ROCmFP4_FAST-GGUF with Ollama:
```
ollama run hf.co/maczzzzzz/Qwen3.6-27B-MTP-ROCmFP4_FAST-GGUF
```

Unsloth Studio

How to use maczzzzzz/Qwen3.6-27B-MTP-ROCmFP4_FAST-GGUF with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for maczzzzzz/Qwen3.6-27B-MTP-ROCmFP4_FAST-GGUF to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for maczzzzzz/Qwen3.6-27B-MTP-ROCmFP4_FAST-GGUF to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for maczzzzzz/Qwen3.6-27B-MTP-ROCmFP4_FAST-GGUF to start chatting

How to use maczzzzzz/Qwen3.6-27B-MTP-ROCmFP4_FAST-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf maczzzzzz/Qwen3.6-27B-MTP-ROCmFP4_FAST-GGUF

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "maczzzzzz/Qwen3.6-27B-MTP-ROCmFP4_FAST-GGUF"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use maczzzzzz/Qwen3.6-27B-MTP-ROCmFP4_FAST-GGUF with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf maczzzzzz/Qwen3.6-27B-MTP-ROCmFP4_FAST-GGUF

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default maczzzzzz/Qwen3.6-27B-MTP-ROCmFP4_FAST-GGUF

Run Hermes

hermes

Atomic Chat new

OpenClaw new

How to use maczzzzzz/Qwen3.6-27B-MTP-ROCmFP4_FAST-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf maczzzzzz/Qwen3.6-27B-MTP-ROCmFP4_FAST-GGUF

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "maczzzzzz/Qwen3.6-27B-MTP-ROCmFP4_FAST-GGUF" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

Docker Model Runner
How to use maczzzzzz/Qwen3.6-27B-MTP-ROCmFP4_FAST-GGUF with Docker Model Runner:
```
docker model run hf.co/maczzzzzz/Qwen3.6-27B-MTP-ROCmFP4_FAST-GGUF
```

Lemonade

How to use maczzzzzz/Qwen3.6-27B-MTP-ROCmFP4_FAST-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull maczzzzzz/Qwen3.6-27B-MTP-ROCmFP4_FAST-GGUF

Run and chat with the model

lemonade run user.Qwen3.6-27B-MTP-ROCmFP4_FAST-GGUF-{{QUANT_TAG}}

List all available models

lemonade list

Qwen3.6-27B-MTP ROCmFP4_FAST — GGUF

ROCmFP4_FAST quant of Qwen/Qwen3.6-27B (Apache 2.0), produced via charlie12345/ROCmFPX. Benchmarked on an RDNA4 RX 9060 XT (16 GB). Context ceiling: 131k at q4_0 KV with flat ~17.2 t/s throughput. Cross-format parity with TQ3_4S (Blackwell) on IFEval: both within 2pp.

MTP speculative decoding is a massive win on this card class: +91% throughput (32.93 t/s) at 64k q4_0 KV. The full model + MTP draft head fits in VRAM at 64k; no PCIe spill.

GSM8K (5-shot, n=50): 0.98 — parity-grade with the Blackwell companion. No server crashes, clean chain-of-thought with calculator annotations. The 0.02 figure in earlier versions was a crash-loop artifact on 16 GB VRAM at full 200-sample load (now corrected).

File

File	Size	Quant	BPW
`Qwen3.6-27B-MTP-ROCmFP4_FAST.gguf`	14.5 GB	Q4_0_ROCMFP4_FAST	~4.25 bpw

NOT a stock llama.cpp quant

ROCmFP4_FAST is a custom weight format from charlie12345/ROCmFPX (Q4_0_ROCMFP4 preset). Stock llama.cpp will exit with unknown quantization at load time. The system_fingerprint of a correctly-served ROCmFPX GGUF is b1-11d76c2 — a different fingerprint means the wrong binary loaded the file.

Scope of these benchmarks — read this first

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

Harness scope is bounded. The numbers come from llama-bench (KV context ladder), lm-eval-harness IFEval n=50, and ctx_scaling_bench (KV precision sweep). That's a regression suite, not a quality benchmark.
Sample sizes are small. Throughput numbers are single-rep on a single GPU. IFEval is n=50. None are powered for multi-seed significance.
No perplexity / wikitext / MMLU. Those are upstream's territory.
Single GPU class (RDNA4 16 GB). All measurements on an RX 9060 XT (gfx1201), ROCm 7.2.3. No Strix unified-memory, no CDNA, no multi-GPU, no Vulkan. Cross-hardware generalization is NOT implied. The companion TQ3_4S quant for Blackwell is in a separate repo.
No human eval. "IFEval parity and flat throughput" is not a quality verdict on this specific quant for every use case.

What this IS good for: a quick signal that the quant (a) loads, (b) runs at sane throughput, (c) doesn't break agent tool-calling, (d) scales predictably with context. 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

Context ceiling (q4_0 KV, AMD RDNA4 RX 9060 XT)

Throughput is flat across the entire tested range. KV precision affects VRAM, not token speed.

| KV type | Ctx | TG tok/s | PP tok/s | VRAM (MiB) | Status | |---|---|---|---:|---:|---:|---| | f16 | 32768 | 17.7 | 102.7 | 16174 | OK | | q4_0 | 65536 | 17.26 | 100.54 | 15282 | OK | | q4_0 | 98304 | 17.22 | 100.11 | 15857 | OK | | q4_0 | 131072 | 17.23 | 99.93 | 16282 | OK | | q8_0 | 65536 | 17.36 | 101.18 | 16236 | OK |

Ceilings: f16 KV caps at 32k, q8_0 at 64k, q4_0 reaches 131k (OOM at 152k — <200 MiB headroom). Asymmetric q8-K + q4-V fails at all ctx sizes.

IFEval n=50 — ROCmFP4_FAST (RDNA4)

Metric	ROCmFP4_FAST (AMD)	TQ3_4S (Blackwell)	Δ
prompt_level_loose	0.32 ± 0.067	0.34 ± 0.068	-0.02
inst_level_loose	0.487	0.474	+0.013
inst_level_strict	0.487	0.461	+0.026

Cross-format parity holds within 2pp on all metrics.

MTP speculative decoding — the RDNA4 win

MTP is a massive throughput win on AMD RDNA4. The full model + 3 GB MTP draft head fits in VRAM at 64k, enabling on-GPU draft verification with no PCIe spill.

Config	Ctx	TG t/s	vs MTP-OFF
MTP-OFF	32k-131k	17.2	baseline
MTP-ON n_max=3 (recommended)	64k	32.93	+91%
MTP-ON n_max=6	32k	28.92	+68% (worse)

Recommendation for AMD RDNA4: use MTP-ON with n_max=3 for contexts ≤64k. For >64k, drop to MTP-OFF (q4_0 KV reaches 131k).

Critical: Do NOT set GGML_HIP_ENABLE_UNIFIED_MEMORY=1 on discrete AMD cards. Charlie's ROCmFPX scripts default to this flag (correct for Strix Halo), but on discrete RDNA4 it moves the ENTIRE model to system RAM (30× regression — 0.97 t/s). The ROCmFP4_FAST quant + draft head fits in 16 GB VRAM at 64k without unified memory.

GSM8K (5-shot, n=50) — corrected finding

Config	Strict
ROCmFP4_FAST (AMD)	0.98 (49/50)

Correction: earlier numbers published 0.02 for this quant. That was a crash-loop artifact: the 16 GB card ran out of VRAM under sustained 200-sample eval load with default --cache-prompt, returning empty responses that scored zero. A clean re-run on 2026-07-06 with --cache-ram 0 --no-cache-prompt and reduced eval batch scored 0.98 (49/50) — authentic chain-of-thought with <<>> calculator annotations and clean #### N answers throughout. The single miss is the classic lemon-tree year-counting ambiguity (model argued 12 years vs. canonical 13 — a defensible reading, not a capability gap). Cross-format GSM8K parity with the Blackwell companion (0.955) is confirmed.

Practical note for users of this quant: on 16 GB AMD cards, the 27B model at ROCmFP4_FAST is at the absolute edge of VRAM. For reliable sustained throughput, serve with --no-cache-prompt --cache-ram 0. If using an eval harness, reduce concurrency or sample count. The quant is high quality — the earlier crash reports were not a model quality issue.

Quick start

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

# Serve (MTP-OFF, direct decode)
./bin/llama-server \
  -m /path/to/Qwen3.6-27B-MTP-ROCmFP4_FAST.gguf \
  --port 8081 \
  -ngl 99 \
  -c 32768 \
  --cache-type-k q4_0 \
  --cache-type-v q4_0 \
  --cache-ram 0 \
  --no-cache-prompt

# Serve (MTP-ON, n_max=3 — recommended for ≤64k ctx)
./bin/llama-server \
  -m /path/to/Qwen3.6-27B-MTP-ROCmFP4_FAST.gguf \
  --port 8081 \
  -ngl 99 \
  -c 65536 \
  --cache-type-k q4_0 \
  --cache-type-v q4_0 \
  --speculative-model m \
  --spec-draft-n-max 3 \
  --cache-ram 0 \
  --no-cache-prompt

Do NOT set GGML_HIP_ENABLE_UNIFIED_MEMORY=1 in the environment on discrete AMD cards.

Reproduce the quant

# From unsloth/Qwen3.6-27B-MTP-GGUF BF16 source (SHA256-validated, single-step)
/path/to/llama-quantize \
  --allow-requantize \
  /path/to/Qwen3.6-27B-MTP-BF16.gguf \
  /path/to/Qwen3.6-27B-MTP-ROCmFP4_FAST.gguf \
  Q4_0_ROCMFP4_FAST

AEON Bench Full Suite — 150 cases (120 scored)

Overall mean: 0.558 (ROCmFP4_FAST) vs 0.525 (TQ3_4S companion)

The full AEON suite covers math, instruction, reasoning, coding, and prose at easy/medium/hard/expert/frontier difficulties. 120 of 150 cases were scored (30 prose cases need a judge endpoint).

Category	ROCmFP4_FAST (AMD)	TQ3_4S (NVIDIA)
Overall Mean	0.558	0.525
math	0.563	0.563
instruction	0.433	0.467
reasoning	0.500	0.467
coding	0.841	0.892
prose	unscored	unscored

See BENCH-aeon-full-suite-240-cases.md for per-case breakdown.

Files in this repo

File	Purpose
`Qwen3.6-27B-MTP-ROCmFP4_FAST.gguf`	The quantized model (LFS-tracked, 14.5 GB)
`README.md`	This file
`BENCH-*.md`	Raw bench data, context-ceiling probes, and cross-format parity reports

What's NOT in this repo (caveats)

Stock llama.cpp will not load this file. ROCmFP4_FAST is a custom weight format unique to charlie12345/ROCmFPX. Use that fork's llama-server.
No CUDA / non-AMD GPU bench. All measurements are RDNA4 (gfx1200). Vulkan path on RDNA4 has a known upstream regression in charlie12345/ROCmFPX.
131k ctx is the practical ceiling on this hardware due to VRAM. 256K requires larger VRAM or smaller model.
No vision/multimodal test. This variant is text-only.
No quality benchmark (perplexity, MMLU). The quant passes IFEval parity and GSM8K parity; whether it's "the best ROCmFP4 quant" needs upstream validation.
16 GB minimum VRAM. Does not fit on smaller cards. The 16 GB card runs it with ~150 MiB headroom at 131k.
GSM8K 0.98 was measured at n=50 5-shot. Initial 0.02 report was a VRAM crash-loop artifact (see corrected finding above). The quant is not broken for reasoning — just tight on 16 GB.

Provenance

Source model: Qwen/Qwen3.6-27B (Apache 2.0)
Intermediate GGUF: unsloth/Qwen3.6-27B-MTP-GGUF (SHA256-verified across nodes, single-step quant)
Quantizer: charlie12345/ROCmFPX commit 5b39566, preset Q4_0_ROCMFP4_FAST
Build hardware: Node B — RDNA4 RX 9060 XT 16 GB (gfx1201), ROCm 7.2.3, NixOS
Bench harnesses: llama-bench (context ladder), lm-eval-harness (IFEval n=50), ctx_scaling_bench (KV precision sweep)

License

The model weights are derived from Qwen/Qwen3.6-27B (Apache 2.0). The ROCmFP4 quant format is provided by charlie12345/ROCmFPX (MIT). This repo is a derivative quant — the Apache 2.0 license of the parent applies to the model weights; the quantizer tooling is separately licensed.