Using AI for Woodworking on the EVO-X-2 with Gemma 4

June 17, 2026

In my previous post I got Rocky Linux 10 running on the GMKtec EVO-X-2 with llama.cpp and full Vulkan GPU offload, serving Qwen3-Coder-Next as a local coding assistant via OpenCode. One of the things I mentioned wanting to explore was using the machine to help plan and research woodworking projects.

This post covers doing exactly that — swapping in Google’s Gemma 4 31B as a second model on the same server and wiring it up in OpenCode for woodworking conversations. The llama.cpp installation and all the kernel and power configuration from the previous guide remain in place; this post only covers the new model.

Why Gemma 4 31B

Gemma 4 is Google DeepMind’s fourth-generation open model family, released in April 2026. The 31B variant is a dense transformer with a 128K token context window and native multimodal support and it can accept both text and images in the same conversation.

That last point is genuinely useful for woodworking. Being able to drop a photo of a rough board, a joinery problem, or a reference piece and discuss it directly in the same context window changes the workflow considerably compared to a text-only model.

At Q4_K_M quantisation the 31B model is approximately 19 GB, which is a fraction of the EVO-X-2’s 96 GB pool — and runs with full GPU offload on the Vulkan backend, leaving ample headroom for a generous context window.

Unsloth provide the quantised GGUF files and benchmark their Dynamic 2.0 variants (UD-*) as higher accuracy at the same bit-width. The UD-Q4_K_XL file is the recommended pick for this hardware. It has better accuracy than standard Q4_K_M with negligible size difference, and well within the memory budget.

Downloading Gemma 4

The hf CLI from the previous guide should already be installed and authenticated. If not, install and log in first:

python3 -m pip install -U huggingface_hub
git config --global credential.helper store
hf auth login

I created a directory for the model on my secondary NVMe drive.

mkdir -p /mnt/data/models/gemma-4-31B-it/
sudo chown $USER:$USER /mnt/data/models/gemma-4-31B-it/

I downloaded the UD-Q4_K_XL quantisation:

hf download unsloth/gemma-4-31B-it-GGUF \
  gemma-4-31B-it-UD-Q4_K_XL.gguf \
  --local-dir /mnt/data/models/gemma-4-31B-it/

Gemma 4 is a multimodal model. To enable image input alongside text, the multimodal projector file is required as well:

hf download unsloth/gemma-4-31B-it-GGUF \
  mmproj-BF16.gguf \
  --local-dir /mnt/data/models/gemma-4-31B-it/

I verified both files were present before proceeding:

$ ls -lh /mnt/data/models/gemma-4-31B-it/
total 20G
-rw-r--r--. 1 howard howard  20G Jun 17 10:42 gemma-4-31B-it-UD-Q4_K_XL.gguf
-rw-r--r--. 1 howard howard 1.6G Jun 17 10:51 mmproj-BF16.gguf

Starting llama-server with Gemma 4

The server flags are largely the same as in the previous guide. Two differences are worth noting:

--mmproj — points to the multimodal projector file to enable image input.
-c 131072 — sets a 128K token context window, matching Gemma 4’s full training context.

The --alias value must match exactly what OpenCode will reference in its configuration:

llama-server \
  --model /mnt/data/models/gemma-4-31B-it/gemma-4-31B-it-UD-Q4_K_XL.gguf \
  --mmproj /mnt/data/models/gemma-4-31B-it/mmproj-BF16.gguf \
  --alias gemma-4-31B-it-UD-Q4_K_XL \
  --host 0.0.0.0 \
  --port 8081 \
  --n-gpu-layers 99 \
  -fa on \
  --parallel 1 \
  -t 32 -tb 32 \
  -ub 2048 \
  -ctk q8_0 -ctv q8_0 \
  --mlock \
  -c 131072

I used port 8081 to avoid conflicting with the Qwen3-Coder-Next server already running on 8080.

I verified the server was healthy:

$ curl http://localhost:8081/health
{"status":"ok"}

I confirmed the model alias was set correctly:

$ curl -s http://localhost:8081/v1/models | python3 -m json.tool

{
    "object": "list",
    "data": [
        {
            "id": "gemma-4-31B-it-UD-Q4_K_XL",
            "aliases": [
                "gemma-4-31B-it-UD-Q4_K_XL"
            ],
            "object": "model",
            "created": 1750111200,
            "owned_by": "llamacpp"
        }
    ]
}

Persisting the Gemma 4 server at boot

Note: if you do not want multiple llama-server instances running at the same time, disable any existing service before enabling this one:

systemctl --user disable --now llama-server.service

I created a second user service unit for the Gemma 4 server alongside the existing Qwen3-Coder-Next unit:

tee ~/.config/systemd/user/llama-server-gemma4.service << 'EOF'
[Unit]
Description=llama.cpp API server (Gemma 4 31B)
After=network-online.target
Wants=network-online.target

[Service]
Type=simple
LimitMEMLOCK=infinity
ExecStart=/usr/local/bin/llama-server \
    --model /mnt/data/models/gemma-4-31B-it/gemma-4-31B-it-UD-Q4_K_XL.gguf \
    --mmproj /mnt/data/models/gemma-4-31B-it/mmproj-BF16.gguf \
    --alias gemma-4-31B-it-UD-Q4_K_XL \
    --host 0.0.0.0 \
    --port 8081 \
    --n-gpu-layers 99 \
    -fa on \
    --parallel 1 \
    -t 32 -tb 32 \
    -ub 2048 \
    -ctk q8_0 -ctv q8_0 \
    --mlock \
    -c 131072
Restart=on-failure
RestartSec=5

[Install]
WantedBy=default.target
EOF

I reloaded the user daemon and enabled the service:

systemctl --user daemon-reload
systemctl --user enable --now llama-server-gemma4.service

I verified it was running:

$ systemctl --user status llama-server-gemma4.service
● llama-server-gemma4.service - llama.cpp API server (Gemma 4 31B)
     Loaded: loaded (/home/howard/.config/systemd/user/llama-server-gemma4.service; enabled; preset: disabled)
     Active: active (running) since ...

The LimitMEMLOCK=infinity override applied in the previous guide to /etc/systemd/system/user@1000.service.d/limits.conf covers all user services, so no additional system configuration was needed here.

Opening the firewall

Port 8081 also needs to be opened in firewalld to allow inbound connections from other machines on the network:

sudo firewall-cmd --permanent --add-port=8081/tcp
sudo firewall-cmd --reload

I verified the rule was active:

$ sudo firewall-cmd --list-ports
8080/tcp 8081/tcp

Configuring OpenCode

OpenCode supports multiple custom providers simultaneously. I added the Gemma 4 server as a second entry alongside the existing evo-x2 provider in ~/.config/opencode/opencode.json.

The model ID in the models map must match the --alias value exactly, and the context limit must match -c:

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "evo-x2": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "EVO-X2 (llama.cpp)",
      "options": {
        "baseURL": "http://<YOUR_SERVER_IP>:8080/v1"
      },
      "models": {
        "Qwen3-Coder-Next": {
          "name": "Qwen3-Coder-Next-Q4_K_M (EVO-X2)",
          "limit": {
            "context": 65536,
            "output": 32768
          }
        }
      }
    },
    "evo-x2-gemma4": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "EVO-X2 Gemma 4 (llama.cpp)",
      "options": {
        "baseURL": "http://<YOUR_SERVER_IP>:8081/v1"
      },
      "models": {
        "gemma-4-31B-it-UD-Q4_K_XL": {
          "name": "Gemma-4-31B-it-UD-Q4_K_XL (EVO-X2)",
          "limit": {
            "context": 131072,
            "output": 65536
          }
        }
      }
    }
  }
}

I ran /models within OpenCode and selected Gemma-4-31B-it-UD-Q4_K_XL (EVO-X2) from the picker to switch over to the Gemma 4 server.

Using Gemma 4 for woodworking

With the model running and wired into OpenCode, I started putting it to use on some actual woodworking problems.

Case Study: Upgrading the Game Shelf

I decided to test Gemma 4’s ability to handle specific dimensions and cultural context by asking it to help me upgrade my Nintendo Switch game shelf. I wanted to add a sliding door that slides outwards, similar to a Japanese sliding door, including a lock.

The Prompt: “I want to upgrade my game shelf (43 cm wide x 40.4 cm tall) using a piece of wood (3 cm thick) to add a sliding door that slides outwards, like a Japanese sliding door. I need a lock as well. Please provide instructions and a shopping list for a Japanese hardware store, including the Japanese terms for the items.”

Gemma 4’s Response: “To create a Japanese-inspired sliding door (similar to a Shoji or Fusuma style) for your game shelf, you’ll want to implement a track system that allows the door to glide horizontally. Given your wood thickness of 3 cm, we will focus on a sturdy, surface-mounted track approach.

Technical Plan

Shelf Dimensions: 43 cm (W) x 40.4 cm (H)
Door Material: 3 cm thick wood
Mechanism: Top and bottom sliding rails

Step-by-Step Instructions

Dimensioning the Door: Cut your wood to approximately 44 cm x 41 cm. This provides a slight overlap (overhang) on the edges to prevent gaps and ensure the ‘outward’ sliding feel.
Installing the Tracks:
- Purchase two U-shaped aluminum channels.
- Mount one channel to the top edge of the shelf and one to the bottom.
- Since you want it to slide ‘outwards’, mount the tracks so they are flush with the front face of the shelf.
Fitting the Door: Insert the door into the channels. If the fit is too tight, sand the edges of the wood slightly.
Adding the Lock: Install a sliding bolt lock on the inner side of the door, securing it to the side wall of the shelf.

Shopping List for a Japanese Hardware Store (ホームセンター - Hōmu Sentā)

To get the right parts, you can use these terms:

English	Japanese (Kanji/Kana)	Romaji
Sliding Rail / U-Channel	アルミ製Uチャンネル / スライドレール	Arumi-sei U-channeru / Suraido reiru
Wood Screws	木ねじ	Kineji
Sliding Bolt Lock	スライドボルト / かんぬき	Suraido boruto / Kannuki
Sandpaper	紙やすり	Kamiyasuri
Wood Glue	木工用ボンド	Mokkō-yō bondo
Right Angle Square	直角定規	Chokkaku jōgi	”

Howard Van Der Wal