Instructions to use michaelw9999/Qwen3.6-27B-NVFP4-MTP-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use michaelw9999/Qwen3.6-27B-NVFP4-MTP-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 michaelw9999/Qwen3.6-27B-NVFP4-MTP-GGUF:NVFP4
# Run inference directly in the terminal:
llama cli -hf michaelw9999/Qwen3.6-27B-NVFP4-MTP-GGUF:NVFP4

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf michaelw9999/Qwen3.6-27B-NVFP4-MTP-GGUF:NVFP4
# Run inference directly in the terminal:
llama cli -hf michaelw9999/Qwen3.6-27B-NVFP4-MTP-GGUF:NVFP4

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 michaelw9999/Qwen3.6-27B-NVFP4-MTP-GGUF:NVFP4
# Run inference directly in the terminal:
./llama-cli -hf michaelw9999/Qwen3.6-27B-NVFP4-MTP-GGUF:NVFP4

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 michaelw9999/Qwen3.6-27B-NVFP4-MTP-GGUF:NVFP4
# Run inference directly in the terminal:
./build/bin/llama-cli -hf michaelw9999/Qwen3.6-27B-NVFP4-MTP-GGUF:NVFP4

Use Docker

docker model run hf.co/michaelw9999/Qwen3.6-27B-NVFP4-MTP-GGUF:NVFP4

LM Studio
Jan
Ollama
How to use michaelw9999/Qwen3.6-27B-NVFP4-MTP-GGUF with Ollama:
```
ollama run hf.co/michaelw9999/Qwen3.6-27B-NVFP4-MTP-GGUF:NVFP4
```

Unsloth Studio

How to use michaelw9999/Qwen3.6-27B-NVFP4-MTP-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 michaelw9999/Qwen3.6-27B-NVFP4-MTP-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 michaelw9999/Qwen3.6-27B-NVFP4-MTP-GGUF to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for michaelw9999/Qwen3.6-27B-NVFP4-MTP-GGUF to start chatting

How to use michaelw9999/Qwen3.6-27B-NVFP4-MTP-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf michaelw9999/Qwen3.6-27B-NVFP4-MTP-GGUF:NVFP4

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": "michaelw9999/Qwen3.6-27B-NVFP4-MTP-GGUF:NVFP4"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

OpenClaw new

How to use michaelw9999/Qwen3.6-27B-NVFP4-MTP-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf michaelw9999/Qwen3.6-27B-NVFP4-MTP-GGUF:NVFP4

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 "michaelw9999/Qwen3.6-27B-NVFP4-MTP-GGUF:NVFP4" \
  --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 michaelw9999/Qwen3.6-27B-NVFP4-MTP-GGUF with Docker Model Runner:
```
docker model run hf.co/michaelw9999/Qwen3.6-27B-NVFP4-MTP-GGUF:NVFP4
```

Lemonade

How to use michaelw9999/Qwen3.6-27B-NVFP4-MTP-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull michaelw9999/Qwen3.6-27B-NVFP4-MTP-GGUF:NVFP4

Run and chat with the model

lemonade run user.Qwen3.6-27B-NVFP4-MTP-GGUF-NVFP4

List all available models

lemonade list

Hermes Agent

How to use michaelw9999/Qwen3.6-27B-NVFP4-MTP-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 michaelw9999/Qwen3.6-27B-NVFP4-MTP-GGUF:NVFP4

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 michaelw9999/Qwen3.6-27B-NVFP4-MTP-GGUF:NVFP4

Run Hermes

hermes

Atomic Chat

Benchmarks: RTX 5090 Desktop — Performance & Intelligence (llama.cpp b9415, 230K ctx)

by raulgdm - opened May 30

Discussion

raulgdm

May 30

Hi Michael, great work on this quantization! I've been running benchmarks with the latest update and wanted to share my results.

Setup

GPU: NVIDIA RTX 5090 (32GB VRAM), desktop
llama.cpp: b9415 (06d26dfdf)
Context: 230K tokens
KV cache: K Q8_0 / V Q5_1
Speculative decoding: MTP + ngram (--spec-type draft-mtp --spec-draft-n-max 3 --spec-default)
CUDA: 13.1
Model file: Qwen3.6-27B-NVFP4-MTP-GGUF.gguf (~15.2 GB)

Performance

All measurements taken via direct API calls to llama-server (/completion and /v1/chat/completions endpoints), generating 512 tokens per run:

Scenario	Gen TPS	MTP Acceptance
Short text (~30 words)	143.4 t/s	84.3%
Medium text (~1000 words)	150.4 t/s	90.1%
Tool calls (function calling)	127.9 t/s	70.7%
Prefill (~2K tokens)	~3,940 t/s	—
Prefill (~8K tokens)	~4,500 t/s	—
Prefill (~32K tokens)	~3,750 t/s	—

VRAM usage: ~30.6 GB / 32.6 GB (94%) — mmproj on GPU.

Intelligence Benchmarks

Compared against other Qwen3.6-27B quantizations tested on the same hardware:

Benchmark	Samples	Q6_K (150K)	NVFP4-Q8_0 (27/05)	NVFP4 1f (29/05)	NVFP4 1f updated (30/05)
GSM8K	51	88.2%	86.3%	88.2%	88.2%
ARC Easy	30	100%	100%	100%	100%
HellaSwag	30	63.3%	66.7%	73.3%	73.3%
PIQA	20	100%	100%	100%	100%
HumanEval	164	85.4%	84.8%	86.0%	85.4%
Average	—	87.9%	88.2%	90.4%	90.4%
PPL (WikiText-2)	313 tokens	—	—	1.13	1.105

All benchmarks used max_tokens=2000 to allow full CoT reasoning. HumanEval uses pass@1 evaluation.

Summary

Intelligence is functionally identical to the previous NVFP4 1f version across all benchmarks
Matches Q6_K on GSM8K and HumanEval, outperforms on HellaSwag (+10pp)
PPL improved slightly (1.13 → 1.105)
MTP acceptance is healthy across all scenarios (70-90%)
Tool calls work well with MTP speculative decoding active
File size increased by ~370 MB (14.87 → 15.24 GB)

Overall a solid quantization — the quality is maintained while the slight PPL improvement is welcome. Thanks for the great work!

michaelw9999

Owner May 31

@raulgdm thank you so much for taking the time to run that eval. I've been working on the quantizer quite a bit and I should start to shift focus to trying to match BF16 on the evaluations more so than just raw numbers. It can take the whole day of running to try to get another 0.01 increase from the quantizer and I see it might not even be worth too much. I came up with a new imatrix dataset instead of running wikitrain, and testing against wikitest; so we'll see where the next one goes. I also should push out a 35B model. Rather than replace the good working model here, I'll probably just put it into this repo as another variant. If there's anything specific you would like to see, just let me know.

michaelw9999

Owner Jun 2

•

edited Jun 2

I ran this identical eval on the NVFP4/MXFP6 blend. It may work even better. It is nearly as fast as all 100% NVFP4 and just a tiny bit bigger but the results are further improved:

Qwen3.6-27B Quantization Benchmarks
Benchmark	Samples	Q6_K (150K)	NVFP4-Q8_0 (27/05)	NVFP4 1f (29/05)	NVFP4 1f updated (30/05)	NVFP4/MXFP6 (w/ repack)
GSM8K	51	88.2%	86.3%	88.2%	88.2%	96.1%
ARC Easy	30	100%	100%	100%	100%	100%
HellaSwag	30	63.3%	66.7%	73.3%	73.3%	86.7%
PIQA	20	100%	100%	100%	100%	100%
HumanEval pass@1	164	85.4%	84.8%	86.0%	85.4%	87.2%
Average	—	87.9%	88.2%	90.4%	90.4%	94.0%
PPL ratio (WikiText-2 full KLD)	—	—	—	1.13	1.105	1.0269

PPL ratio = Mean PPL(Q)/PPL(base) = 1.026890 from the full KLD/PPL run.

FallingSnowStar

Jun 5

@raulgdm What command are you using to get those numbers? I haven't seen anything higher than 75 tk/s.

labhraighlep

Jul 14

•

edited Jul 14

FYI only because I thought this model was interesting enough to include in my own testing on a 5090. I'm loving the lower VRAM usage with the resulting overhead available for KV Cache. I'm getting 123-124 tokens/sec using the following

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment