Update local LLMs documentation to feature Qwen3-Coder-30B-A3B and Devstral Small 2

[neubig]

Summary

This PR updates the local LLMs documentation to feature the two most recent and powerful models: Qwen3-Coder-30B-A3B-Instruct and Devstral Small 2 (24B), removing references to older models.

Changes Made

1. Updated Model Information

• ✅ Featured Qwen3-Coder-30B-A3B-Instruct (replacing older Qwen2.5-Coder references)
• ✅ Featured Devstral-Small-2-24B-Instruct-2512 (replacing Devstral Small 2505)
• ✅ Removed references to OpenHands LM 32B v0.1 and older Devstral versions

2. Hardware Requirements

• ✅ Added accurate VRAM requirements based on real-world testing
  • Qwen3-Coder: 12GB VRAM minimum (tested on RTX 3060 12GB + 64GB RAM)
  • Devstral Small 2: 16GB VRAM minimum
• ✅ Included Mac Apple Silicon requirements for both models

3. Context Window Recommendations

• ✅ Added guidance for context window settings (22k minimum for Qwen3-Coder, 32k recommended)

4. Updated All Deployment Methods

• ✅ LM Studio: Updated model search names and configuration examples
• ✅ Ollama: Corrected pull commands to qwen3-coder:30b and devstral-small-2
• ✅ SGLang: Updated model paths and served-model-name parameters
• ✅ vLLM: Updated model paths and configurations
• ✅ HuggingFace CLI: Corrected repository paths for model downloads

5. Configuration Examples

• ✅ Updated Custom Model identifiers for all backends
• ✅ Ensured consistency across LM Studio, Ollama, SGLang, and vLLM examples

Verification

All model names and commands have been verified against:

• Official HuggingFace repositories: Qwen3-Coder-30B-A3B-Instruct and Devstral-Small-2-24B-Instruct-2512
• Ollama library listings for correct model tags
• User feedback from community testing (RTX 3060 compatibility)

Files Changed

• openhands/usage/llms/local-llms.mdx

Context

Based on community feedback, these two models represent the current state-of-the-art for local coding assistance with OpenHands:

• Qwen3-Coder-30B-A3B shows excellent performance even on consumer hardware (12GB VRAM)
• Devstral Small 2 achieves remarkable results on SWE-bench and agentic coding tasks

Both models are significantly more capable than the previously documented options.

@neubig can click here to continue refining the PR