opencode-gemma4-setup

Local autonomous coding agent stack for AMD RX 6700 XT (RDNA2) on Windows. OpenCode + LM Studio (Vulkan) + Gemma 4 E4B Q6_K + oh-my-opencode plugin. Runs at ~30 tok/s on 12GB VRAM. Free, private, GPU-accelerated.

Why this exists

If you have an AMD RDNA2 GPU (RX 6600/6700/6800 series) on Windows and want a Claude Code-style autonomous coding agent for free, locally, with GPU acceleration, you'll hit a wall:

• Ollama Vulkan + Gemma 4 = broken upstream (issues #15328, #15248, #15285). Q4 outputs ktktktkt..., Q8 outputs garbage. CPU-only works but ~5 tok/s.
• vLLM ROCm is unofficial for gfx1031. Build is hell.
• llama.cpp Vulkan direct build works but needs cmake/MSVC/Vulkan SDK toolchain (30-60 min).
• LM Studio (Vulkan) uses a different ggml build → Gemma 4 works perfectly, GPU 100%.

This repo packages the validated config so you skip the 6+ hour debugging session.

What's included

File	Purpose
configs/opencode.json	Global OpenCode config — LM Studio + Ollama providers, MCP servers, default model
configs/AGENTS.md	OpenCode global guide — Gemma 4 environment hints, camelCase tool params
configs/oh-my-openagent.json	omo plugin config (Claude max20 + opencode/gpt-5-nano agent assignments)
configs/Modelfile.gemma4-fast	Ollama Modelfile (legacy, unused with LM Studio path)
INSTALL.md	Step-by-step install guide
TROUBLESHOOTING.md	Vulkan + Gemma 4 issue notes, model selection matrix

Quick benchmark (validated)

Test	Result
Gemma 4 E4B Q6_K load	5.0s on RX 6700 XT
Coding (Python fib)	clean code + docstring + edge cases
Tool calling	native, finish_reason: tool_calls
GPU usage	3.8GB VRAM (Q6_K, 8K context)
Inference speed	~30 tok/s

Stack

┌─────────────────────────────────────┐
│  OpenCode v1.4.11+ (TUI/CLI)        │
│  + oh-my-opencode plugin            │
└──────────────┬──────────────────────┘
               │ OpenAI-compat HTTP
               ▼
┌─────────────────────────────────────┐
│  LM Studio (port 1234)              │
│  Vulkan backend                     │
│  google/gemma-4-e4b Q6_K (6.33GB)   │
└──────────────┬──────────────────────┘
               │ Vulkan API
               ▼
┌─────────────────────────────────────┐
│  AMD RX 6700 XT (RDNA2 / gfx1031)   │
│  12GB VRAM, Windows 11              │
└─────────────────────────────────────┘

Model selection matrix (validated)

Workload	Model	Provider
General coding	lmstudio/google/gemma-4-e4b	LM Studio (Q6_K, GPU) ← default
Smarter cloud work	claude-opus-4-6	omo + Claude Pro/Max20
Korean-heavy text	exaone-deep:7.8b	Ollama (CPU/GPU)
Larger thinking model	qwen3.5:9b	Ollama (GPU works for non-Gemma)
Multimodal (vision/audio)	Gemma 4 E4B only	LM Studio

Install

See INSTALL.md for step-by-step.

TL;DR:

# 1. Install OpenCode
bun install -g opencode-ai

# 2. Install LM Studio (winget) — first-time GUI launch required
winget install -e --id ElementLabs.LMStudio
# Open LM Studio once to init ~/.lmstudio

# 3. Download Gemma 4 E4B Q6 in LM Studio (Hub UI or `lms get`)

# 4. Copy configs
mkdir -p ~/.config/opencode
cp configs/opencode.json ~/.config/opencode/
cp configs/AGENTS.md ~/.config/opencode/
# Edit opencode.json: replace ${HOME}/your/project/root with your actual project path

# 5. Install oh-my-opencode plugin
bunx oh-my-opencode install --no-tui --claude=<yes|max20|no> --gemini=no --copilot=no

# 6. Start LM Studio server
lms server start
lms load google/gemma-4-e4b --gpu max --context-length 8192

# 7. Run
opencode

Troubleshooting

See TROUBLESHOOTING.md. Most common: garbage output → you're using Ollama Vulkan + Gemma 4. Switch to LM Studio.

License

MIT — see LICENSE.

Credits

• OpenCode by SST
• oh-my-opencode by code-yeongyu
• LM Studio by Element Labs
• Gemma 4 by Google DeepMind