Run Muse Glimmer locally with llama.cpp, DFlash, and Pi

Published August 22, 2026

a computer screen with a bunch of code on it

If you have an RTX 3090, 4090, or 5090 sitting under your desk, you can run a 30B coding model locally today. No cloud bill, no code sharing with third-party services, no per-token pricing.

Muse Glimmer is the model worth trying right now. It’s getting compared to Qwen’s 27B-class models in the local AI community, and in many cases it’s performing better for coding and agentic workflows. This guide walks through the full setup: downloading Muse Glimmer, serving it with llama.cpp, accelerating it with DFlash speculative decoding, and connecting it to the Pi coding agent so it can build, test, and debug projects directly from your terminal.

Table of Contents

‍ Step 1: Download Muse Glimmer

You need two files from Hugging Face: the main model and the DFlash drafter. First, install the Hugging Face CLI:

curl -LsSf https://hf.co/cli/install.sh | bash
echo 'export PATH="/root/.local/bin:$PATH"' >> ~/.bashrc
source ~/.bashrc

hf auth login
mkdir -p /workspace/muse-glimmer

Download the main model (16.8 GB):

hf download meta-models/Muse-Glimmer-30B-GGUF 
  muse-glimmer-30B-kquant-17gb.gguf 
  --local-dir /workspace/muse-glimmer

Download the DFlash drafter (1.63 GB):

hf download meta-models/Muse-Glimmer-30B-GGUF 
  dflash-kquant.gguf 
  --local-dir /workspace/muse-glimmer

Both files land in /workspace/muse-glimmer.

Step 2: Install and Build llama.cpp with CUDA

Clone the repo and build with CUDA support:

cd /workspace
git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
git pull origin master
cmake -B build -DGGML_CUDA=ON -DCMAKE_BUILD_TYPE=Release
cmake --build build --config Release -j$(nproc)
ln -sf "$(pwd)/build/bin/llama-server" /root/.local/bin/llama-server

Once built, start the server with DFlash speculative decoding enabled:

llama-server 
  -m /workspace/muse-glimmer/muse-glimmer-30B-kquant-17gb.gguf 
  -md /workspace/muse-glimmer/dflash-kquant.gguf 
  --spec-type draft-dflash 
  --spec-draft-n-max 15 
  -ngl all 
  --spec-draft-ngl all 
  -fa on 
  --ctx-size 16384 
  --alias muse 
  --host 0.0.0.0 
  --port 8080 
  --jinja

llama.cpp loads both the main model and the DFlash drafter onto the GPU. Speculative decoding lets the smaller drafter propose tokens ahead of the main model, which is where the speed gains come from.

⏱️ Step 3: Test Speed in the Built-in Web UI

Once the server is running, open the llama.cpp web interface at http://localhost:8080/ to run a quick sanity check.

In initial testing, the model hit around 46 tokens per second. During longer coding tasks, that climbed to around 127 tokens per second, which makes agentic workflows feel genuinely responsive.

One honest note: the results were mixed on HTML tasks. Muse Glimmer produced an HTML game on request, but it didn’t work well. For producing working HTML apps and games, Qwen3.8-27B was noticeably better in the tester’s experience.

️ Step 4: Install Pi and Connect to llama.cpp

Pi is the coding agent that connects to your local llama.cpp server and handles the agentic loop. Install it:

curl -fsSL https://pi.dev/install.sh | sh

Then install Hugging Face’s llama.cpp extension for Pi:

pi install git:github.com/huggingface/pi-llama

Restart your terminal after installation. The pi-llama extension automatically connects to http://localhost:8080/v1 and detects whatever models llama.cpp is serving. You don’t need to edit models.json manually.

Step 5: Select Muse Glimmer Inside Pi

Woman meditating on yoga mat with phone and drink

Create a project directory and launch Pi:

mkdir -p /workspace/glimmer-test
cd /workspace/glimmer-test
pi

Inside Pi, run /model, search for llama-cpp, then select muse. Muse Glimmer is now available through Pi’s llama-cpp provider.

Step 6: Run a Full Agentic Coding Task

This is where Muse Glimmer earns its place. Give it a complete project to build from scratch, including tests and documentation, and tell it not to ask you to run anything manually. Here’s the prompt used in testing:

Build a complete Python task management API from scratch using FastAPI. Create a clean project structure. Add endpoints to create, list, update, and delete tasks. Use SQLite for persistence. Add input validation and error handling. Add pytest tests for all endpoints. Create requirements.txt and README.md. Run the tests yourself. Fix any errors and rerun the tests until everything passes. Do not ask me to create files or run commands for you. Build and test the complete project yourself.

Muse Glimmer completed the project in around 2 minutes. To verify locally:

pip install -r requirements.txt
uvicorn app.main:app --reload

Open the API docs at http://localhost:8000/docs. You can also hand Muse Glimmer the task of testing the complete API itself and generating a final report, rather than checking each endpoint manually.

For agentic coding specifically, this is where the model impressed most. It handled multi-step tasks cleanly and took only a few seconds to identify and fix issues during debugging.

Honest Verdict

Muse Glimmer is a solid local coding model, especially given that Meta provides the official GGUF files and recommended configuration. Setup is straightforward. The DFlash speculative decoding makes token speeds genuinely usable for agentic loops.

The rough edges are real: HTML generation underperformed compared to Qwen3.8-27B. But for FastAPI projects, multi-file Python work, and terminal-driven agentic tasks, it holds up well.

If you have an RTX 3090, 4090, or 5090, it’s becoming harder to justify paying per-request for AI coding help or sending your code to third-party services. Local models are already getting close to hosted options like GLM-5.2, and the next few iterations of both the model and the surrounding tooling should close that gap further.

Cagri Sarigoz( Founder )

I’m founder of BizStack at Cagri Sarigoz LLC and a passionate advocate for entrepreneurs.

With over 14 years in tech, marketing, and AI, including my role as Head of SEO at CitizenShipper and co-founder of TaleBot at Intale AI, I’m dedicated to sharing genuine, useful product insights and tips.

At BizStack, I aim to cut through the digital noise to provide clear, actionable advice.

And more than all else, I’m a father to a (always) little girl and a husband.

Contact me at [email protected] for assistance.