See also: Go development
This is mainly about developing with Go and AI code assistants. It includes tooling and configuration options.
AI coding agents
This focuses on affordable and open source models:
- Deep Seek (V4 Pro Max)
- Kimi (K2.6/K2.7)
- Qwen 2.5 Coder 7B
Go libraries
- go-deepseek/deepseek-go
- aicli
- dsk
VSCodium extensions
- Cline (AI assistant that can support Deepseek, Kimi, Ollama, etc.)
- Kimi Code for VSCODE
- Kimi Code CLI
curl -fsSL https://code.kimi.com/kimi-code/install.sh | bash`
Development stacks
- For Kimi which is cheapest for development, Kimi Code CLI/VSCODE, Moonshot API.
- For local (free) Qwen 2.5 Coder 7B on Ollama with Cline VSCODE
Ollama + Qwen + Cline
- Install Ollama (the easiest way): Download from ollama.com. It handles GPU acceleration automatically on your Radeon (via Vulkan or ROCm).
- Pull the model: Open your terminal and run: ollama run qwen2.5-coder:7b
- Configure Cline:
- Open Cline in VSCodium.
- Go to its settings (the gear icon).
- Under "API Provider," select "Ollama".
- Set the base URL to http://localhost:11434.
- In the "Model ID" field, type qwen2.5-coder:7b.
- Test it: Give Cline a simple Go task (e.g., "write a function to reverse a linked list"). Check the token-per-second speed in the output window. If it's below 8 tokens/sec, switch to the qwen2.5-coder:7b (if you tried 14B) or use the Q3_K_M version to shrink memory usage.
Performance Tips
- Close other apps: The 680M is sensitive to system memory bandwidth. Browsers with many tabs will compete for RAM and slow down inference.
- On Linux: Use the Radeon open-source driver with Vulkan; performance is often better than Windows for LLM inference.
- Adjust context length: In Cline, keep the context window small (~8k-16k) to prevent out-of-memory errors. Ollama defaults to 2048, but you can increase it cautiously.
- Bottom line: This setup is 100% viable. You won't get GPT-4 speeds, but Qwen 2.5 Coder 7B on your 680M will be fast enough for Cline's agentic workflows.