I'd had my internet fail a few times recently leaving me stranded without a coding agent, and so when I saw the "Gemma 4 now runs 2x faster with MTP" Multi-Token Prediction update for Gemma 4 I decided to have a go at getting it running.
I wanted a local coding agent setup that:
- was fast enough to actually use on my Mac
- worked through an OpenAI compatible API (so I could use it in other tools)
- and preferably could handle screenshots/images when needed, so I can feed it screenshots of what it has made.
And I did! This video is realtime. And shows the agent responding at a perfectly usable speed.
After a bit of testing the final setup I ended up with is:
- llama.cpp built with Metal on macOS
- Gemma 4 26B-A4B in GGUF format
- A Q8 MTP draft model for speculative decoding
- The Gemma 4 multimodal projector
- Pi as the terminal coding agent
This was tested on an Apple M1 Max with 64 GB unified memory, running macOS 15.7.7.
The main model is: gemma-4-26B-A4B-it-UD-Q4_K_XL.gguf.
Link on Huggingface: models/unsloth-gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_XL.gguf
That file is about 16 GB. With the MTP draft head and multimodal projector the model folder is about 17 GB.
The benchmark prompt was:
Write a compact Python function that parses a unified diff and returns the changed file paths. Then explain two edge cases.
Each benchmark generated about 128 tokens.
First I ran the main model directly through llama.cpp with Metal acceleration:
repos/llama.cpp/build/bin/llama-cli \
-m models/unsloth-gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_XL.gguf \
-ngl 999 \
-fa on \
-c 4096 \
-n 128
Result:
| Setup | Prompt tok/s | Generation tok/s |
|---|---|---|
| Gemma 4 26B-A4B Q4, llama.cpp Metal | 298.0 | 58.2 |
58 tokens/second is not fast, but is usable, but for coding-agent work you want it to be as fast as possible, especially when the agent is making many tool calls.
Gemma 4 now has the MTP draft model available:
MTP/gemma-4-26B-A4B-it-Q8_0-MTP.gguf
This can be loaded by llama.cpp as a speculative draft model:
repos/llama.cpp/build/bin/llama-cli \
-m models/unsloth-gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_XL.gguf \
--model-draft models/unsloth-gemma-4-26B-A4B-it-GGUF/MTP/gemma-4-26B-A4B-it-Q8_0-MTP.gguf \
--spec-type draft-mtp \
--spec-draft-n-max 3 \
-ngl 999 \
-fa on \
-c 4096 \
-n 128
The first run with MTP came in at 69.2 tokens/second using 4 draft tokens. However, Unsloth's guide on How to Run MTP Models includes this note:
"We found --spec-draft-n-max 2 is the best starting point however, do not assume 2 is optimal, as performance is hardware-dependent. Try any value from 1 through 6 and use whichever is fastest for your system."
After sweeping --spec-draft-n-max, the best result was 72.2 tokens/second with 3 draft tokens.
| Setup | Prompt tok/s | Generation tok/s | Speedup |
|---|---|---|---|
| Main model only | 298.0 | 58.2 | 1.00x |
| Main model + Q8 MTP draft | 295.6 | 72.2 | 1.24x |
The useful part is that prompt processing stayed basically the same, while generation improved by about 24%.
I tested --spec-draft-n-max values from 1 to 6.
--spec-draft-n-max |
Prompt tok/s | Generation tok/s |
|---|---|---|
| 1 | 295.5 | 68.4 |
| 2 | 299.1 | 72.0 |
| 3 | 295.6 | 72.2 |
| 4 | 297.3 | 70.7 |
| 5 | 297.9 | 63.7 |
| 6 | 296.3 | 61.2 |
On my M1 Max machine, 3 was the fastest, with 2 close enough that either would be fine. Values above that got slower.
I also tested MLX models through mlx-lm, to find out which is the faster way to run the model on a Mac, llama.cpp or mlx.
| Runtime | Model | Generation tok/s |
|---|---|---|
| llama.cpp Metal + MTP | Unsloth GGUF Q4 + Q8 MTP | 72.2 |
| llama.cpp Metal | Unsloth GGUF Q4 | 58.2 |
| MLX-LM | Unsloth UD MLX 4-bit | 45.8 |
| MLX-LM | mlx-community 4-bit | 43.9 |
| MLX-LM | mlx-community OptiQ 4-bit | 38.1 |
I thought MLX (being optimised for the Mac) would be fastest.
However, for this specific setup, llama.cpp was faster than MLX, and llama.cpp with MTP was clearly the best option.
I guess all the effort and tweaking which has gone into llama.cpp over time means it quite well optimised fr macOS despite being cross platform.
I also tried Gemma 4 MTP through gemma-4-swift-mlx, but the tested 26B 4-bit MLX checkpoints did not match the loader's expected weight keys, and I already had the previous MLX tests, so moved on rather than redownload new models and try to tweak things to match.
For Pi, I also wanted to be able to attach screenshots. The local model entry I setup for it originally declared the model as text-only:
That meant Pi did not send image tool output through to the model properly.
The llama.cpp server also needs the Gemma 4 multimodal projector in order for the multi-modal part to work (only the 12B is natively multi-modal):
When loaded with --mmproj, llama.cpp advertises multimodal support, and Pi can send images.
I re-ran the text benchmark with the projector loaded, just to check it didn't change the speed:
| Setup | Projector | Prompt tok/s | Generation tok/s |
|---|---|---|---|
| llama.cpp Metal + MTP | none | 120.3 | 71.4 |
| llama.cpp Metal + MTP | mmproj-BF16.gguf |
297.4 | 72.2 |
The final run with the projector did not show a text-generation slowdown.
Now for setup instructions:
Install dependencies:
brew install cmake git tmux [email protected]
Clone and build llama.cpp:
mkdir -p ~/Developer/ML-Models/Gemma4/repos
cd ~/Developer/ML-Models/Gemma4
git clone https://github.com/ggml-org/llama.cpp repos/llama.cpp
cd repos/llama.cpp
cmake -B build \
-DCMAKE_BUILD_TYPE=Release \
-DGGML_METAL=ON \
-DGGML_ACCELERATE=ON
cmake --build build --config Release -j
The build I tested had:
GGML_METAL=ON
GGML_ACCELERATE=ON
GGML_BLAS=ON
GGML_BLAS_VENDOR=Apple
Create a Python environment:
cd ~/Developer/ML-Models/Gemma4
python3.11 -m venv .venv
source .venv/bin/activate
pip install -U huggingface_hub hf_xet
Download the files:
mkdir -p models/unsloth-gemma-4-26B-A4B-it-GGUF
huggingface-cli download unsloth/gemma-4-26B-A4B-it-GGUF \
gemma-4-26B-A4B-it-UD-Q4_K_XL.gguf \
mmproj-BF16.gguf \
MTP/gemma-4-26B-A4B-it-Q8_0-MTP.gguf \
--local-dir models/unsloth-gemma-4-26B-A4B-it-GGUF
You should end up with:
models/unsloth-gemma-4-26B-A4B-it-GGUF/
gemma-4-26B-A4B-it-UD-Q4_K_XL.gguf
mmproj-BF16.gguf
MTP/gemma-4-26B-A4B-it-Q8_0-MTP.gguf
This is the final server command:
repos/llama.cpp/build/bin/llama-server \
-m models/unsloth-gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_XL.gguf \
--model-draft models/unsloth-gemma-4-26B-A4B-it-GGUF/MTP/gemma-4-26B-A4B-it-Q8_0-MTP.gguf \
--mmproj models/unsloth-gemma-4-26B-A4B-it-GGUF/mmproj-BF16.gguf \
--spec-type draft-mtp \
--spec-draft-n-max 3 \
-ngl 999 \
-fa on \
-c 65536 \
--parallel 1 \
--host 127.0.0.1 \
--port 8080
The OpenAI-compatible endpoint is:
I used a small start_server.sh wrapper so it runs inside tmux:
#!/usr/bin/env bash
set -euo pipefail
ROOT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
SESSION_NAME="${SESSION_NAME:-gemma4-server}"
HOST="${HOST:-127.0.0.1}"
PORT="${PORT:-8080}"
CTX_SIZE="${CTX_SIZE:-65536}"
PARALLEL="${PARALLEL:-1}"
LLAMA_SERVER="$ROOT_DIR/repos/llama.cpp/build/bin/llama-server"
MODEL="$ROOT_DIR/models/unsloth-gemma-4-26B-A4B-it-GGUF/gemma-4-26B-A4B-it-UD-Q4_K_XL.gguf"
DRAFT_MODEL="$ROOT_DIR/models/unsloth-gemma-4-26B-A4B-it-GGUF/MTP/gemma-4-26B-A4B-it-Q8_0-MTP.gguf"
MMPROJ="$ROOT_DIR/models/unsloth-gemma-4-26B-A4B-it-GGUF/mmproj-BF16.gguf"
LOG_FILE="$ROOT_DIR/logs/llama-server-mtp.log"
mkdir -p "$ROOT_DIR/logs"
tmux new-session -d -s "$SESSION_NAME" -c "$ROOT_DIR" \
"$LLAMA_SERVER \
-m '$MODEL' \
--model-draft '$DRAFT_MODEL' \
--mmproj '$MMPROJ' \
--spec-type draft-mtp \
--spec-draft-n-max 3 \
-ngl 999 \
-fa on \
-c '$CTX_SIZE' \
--parallel '$PARALLEL' \
--host '$HOST' \
--port '$PORT' \
2>&1 | tee -a '$LOG_FILE'"
Start it:
chmod +x start_server.sh
./start_server.sh
Check that the server is running:
curl http://127.0.0.1:8080/v1/models
Pi reads model providers from:
Add a local provider:
{
"providers": {
"gemma4-local": {
"name": "Gemma 4 Local",
"baseUrl": "http://127.0.0.1:8080/v1",
"api": "openai-completions",
"apiKey": "local",
"authHeader": false,
"compat": {
"supportsDeveloperRole": false,
"supportsReasoningEffort": false
},
"models": [
{
"id": "gemma-4-26B-A4B-it-UD-Q4_K_XL.gguf",
"name": "Gemma 4 26B-A4B Q4 + MTP",
"reasoning": false,
"input": ["text", "image"],
"contextWindow": 65536,
"maxTokens": 8192,
"cost": {
"input": 0,
"output": 0,
"cacheRead": 0,
"cacheWrite": 0
}
}
]
}
}
}
The important pieces are:
baseUrlpoints to the llama.cpp OpenAI-compatible server.apiisopenai-completions.authHeaderisfalse, because this is a local server.inputincludes bothtextandimage, otherwise Pi treats it as text-only.
Optionally make it the default in:
~/.pi/agent/settings.json
{
"defaultProvider": "gemma4-local",
"defaultModel": "gemma-4-26B-A4B-it-UD-Q4_K_XL.gguf",
"defaultThinkingLevel": "minimal"
}
Then check Pi can see it:
pi --offline --list-models gemma
Expected:
provider model context max-out thinking images
gemma4-local gemma-4-26B-A4B-it-UD-Q4_K_XL.gguf 65.5K 8.2K no yes
Run Pi using the local model:
pi --provider gemma4-local --model gemma-4-26B-A4B-it-UD-Q4_K_XL.gguf
Or use non-interactive mode:
pi -p --provider gemma4-local --model gemma-4-26B-A4B-it-UD-Q4_K_XL.gguf \
"Explain what this repository does"
For screenshots:
pi -p @"/path/to/screenshot.png" "Describe this image and point out anything relevant to the UI"
The final local coding-agent stack was:
| Layer | Choice |
|---|---|
| Inference runtime | llama.cpp |
| macOS acceleration | Metal + Accelerate |
| Main model | gemma-4-26B-A4B-it-UD-Q4_K_XL.gguf |
| Draft model | gemma-4-26B-A4B-it-Q8_0-MTP.gguf |
| MTP setting | --spec-draft-n-max 3 |
| Multimodal projector | mmproj-BF16.gguf |
| Server | llama-server on 127.0.0.1:8080 |
| API | OpenAI-compatible /v1 |
| Coding agent | Pi |
| Pi model input | ["text", "image"] |
The main conclusion was that the MTP draft model is worth using. On this machine it took Gemma 4 from 58.2 tokens/second to 72.2 tokens/second, while keeping the setup simple enough to run as a local OpenAI-compatible server.
P.S: Some suggested using Qwen3.6 35B-A3B instead of Gemma 4 26B-A4B. According to the benchmarks I can find, Qwen is a much better coding agent than Gemma 4.
However, it is also slower. Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf + unsloth-Qwen3.6-35B-A3B-MTP-GGUF + mmproj-BF16.gguf results in 55 tk/s, instead of 72 tk/s. Which is quite significant when you are sitting waiting for it.
Download the models:
mkdir -p models/unsloth-Qwen3.6-35B-A3B-MTP-GGUF
huggingface-cli download unsloth/Qwen3.6-35B-A3B-MTP-GGUF \
Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf \
mmproj-BF16.gguf \
--local-dir models/unsloth-Qwen3.6-35B-A3B-MTP-GGUF
Start the server:
LLAMA_SERVER=/Users/kylehowells/Developer/ML-Models/Gemma4/repos/llama.cpp/build/bin/llama-server
$LLAMA_SERVER \
-m models/unsloth-Qwen3.6-35B-A3B-MTP-GGUF/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf \
--mmproj models/unsloth-Qwen3.6-35B-A3B-MTP-GGUF/mmproj-BF16.gguf \
--spec-type draft-mtp \
--spec-draft-n-max 3 \
-ngl 999 \
-fa on \
-c 65536 \
--parallel 1 \
--host 127.0.0.1 \
--port 8081
Pi Config:
{
"providers": {
"qwen36-local": {
"name": "Qwen3.6 Local",
"baseUrl": "http://127.0.0.1:8081/v1",
"api": "openai-completions",
"apiKey": "local",
"authHeader": false,
"compat": {
"supportsDeveloperRole": false,
"supportsReasoningEffort": false
},
"models": [
{
"id": "Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf",
"name": "Qwen3.6 35B-A3B Q4 + MTP",
"reasoning": true,
"input": ["text", "image"],
"contextWindow": 65536,
"maxTokens": 8192,
"cost": {
"input": 0,
"output": 0,
"cacheRead": 0,
"cacheWrite": 0
}
}
]
}
}
}