Instructions to use sjakek/gemma4-12b-mtp-assistant with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use sjakek/gemma4-12b-mtp-assistant 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 sjakek/gemma4-12b-mtp-assistant:BF16
# Run inference directly in the terminal:
llama cli -hf sjakek/gemma4-12b-mtp-assistant:BF16

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf sjakek/gemma4-12b-mtp-assistant:BF16
# Run inference directly in the terminal:
llama cli -hf sjakek/gemma4-12b-mtp-assistant:BF16

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 sjakek/gemma4-12b-mtp-assistant:BF16
# Run inference directly in the terminal:
./llama-cli -hf sjakek/gemma4-12b-mtp-assistant:BF16

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 sjakek/gemma4-12b-mtp-assistant:BF16
# Run inference directly in the terminal:
./build/bin/llama-cli -hf sjakek/gemma4-12b-mtp-assistant:BF16

Use Docker

docker model run hf.co/sjakek/gemma4-12b-mtp-assistant:BF16

LM Studio
Jan

vLLM

How to use sjakek/gemma4-12b-mtp-assistant with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "sjakek/gemma4-12b-mtp-assistant"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "sjakek/gemma4-12b-mtp-assistant",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/sjakek/gemma4-12b-mtp-assistant:BF16

Ollama
How to use sjakek/gemma4-12b-mtp-assistant with Ollama:
```
ollama run hf.co/sjakek/gemma4-12b-mtp-assistant:BF16
```

Unsloth Studio

How to use sjakek/gemma4-12b-mtp-assistant 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 sjakek/gemma4-12b-mtp-assistant 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 sjakek/gemma4-12b-mtp-assistant to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for sjakek/gemma4-12b-mtp-assistant to start chatting

Docker Model Runner
How to use sjakek/gemma4-12b-mtp-assistant with Docker Model Runner:
```
docker model run hf.co/sjakek/gemma4-12b-mtp-assistant:BF16
```

Lemonade

How to use sjakek/gemma4-12b-mtp-assistant with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull sjakek/gemma4-12b-mtp-assistant:BF16

Run and chat with the model

lemonade run user.gemma4-12b-mtp-assistant-BF16

List all available models

lemonade list

Atomic Chat

sjakek commited on Jun 3

Commit

e2207f2

verified ·

1 Parent(s): b889cfa

Document parallel-3 MTP benchmark results

Browse files

Files changed (1) hide show

README.md +10 -0

README.md CHANGED Viewed

@@ -90,6 +90,16 @@ Relevant retained logs:
 - `logs/server-final-q8-mtp-canonical.log`
 - `logs/openai-chat-final-q8-canonical.json`
 ## Checksums
 ```text

 - `logs/server-final-q8-mtp-canonical.log`
 - `logs/openai-chat-final-q8-canonical.json`
+## Benchmarks
+Additional local benchmark reports are included in this repository:
+- `benchmarks/gemma4_q4xl_mtp_2k_64k`: single-stream Q4_K_XL target-only vs Q8 MTP, draft max 3, temperature 0.
+- `benchmarks/gemma4_q4xl_mtp_2k_64k_draft2_temp06_topk1`: single-stream Q4_K_XL target-only vs Q8 MTP, draft max 2, drafter top-k 1, accepter temperature 0.6.
+- `benchmarks/gemma4_q4xl_mtp_2k_64k_draft2_temp06_topk1_parallel3`: three concurrent 2K-in/2K-out streams, `--parallel 3`, draft max 2, drafter top-k 1, accepter temperature 0.6. This was the first local run where MTP beat target-only on aggregate generation throughput:
+  - coding: 47.14 tok/s target-only vs 50.57 tok/s MTP
+  - general: 48.42 tok/s target-only vs 50.60 tok/s MTP
 ## Checksums
 ```text