Qwen Models with Claude Code on 36gb vram - insights

A developer's hands-on benchmark of two Alibaba Qwen models running locally with Claude Code has provided valuable, practical data for the AI community. Testing the Qwen3-Coder-Next-80B and Qwen3.5-35B models using Unsloth's GGUF quantizations, both models successfully loaded into a combined 36GB of VRAM from an RTX 3090 and RTX 5070, handling a context length of approximately 132,000 tokens. The user noted the 35B model could potentially run a higher 5 or 6-bit quantization with the available memory.

The key finding was a stark difference in stability and performance. The larger Qwen3-Coder-Next-80B model was deemed "superior in all aspects," reliably completing coding tasks without errors in tool calls. In contrast, the Qwen3.5-35B model consistently failed by stopping in the middle of jobs within Claude Code, requiring manual intervention via the `/execute-plan` command from Superpowers to continue. Despite trying suggested parameters and updating to the latest Unsloth GGUF release to address a known bug, the 35B model's performance remained unsatisfactory.

For developers using Claude Code as a local AI coding agent, the 80B model's performance was comparable to using Anthropic's cloud-based Claude 3.5 Sonnet (the previous version) in terms of speed and, more importantly, reliability. This real-world test challenges the simple assumption that a smaller model is always preferable for local deployment, highlighting that stability and correct tool use can be more critical than raw parameter count for practical applications.