100% Free Claude Code | Run Claude Code with Local LLM with Ollama and Qwen 3.5

Running Claude Code locally with an Ollama-backed Qwen model can deliver practical coding assistance—especially when the task is narrowly scoped to specific files—without relying on Anthropic’s hosted, quota-based API. The setup is straightforward: install Claude Code, then launch it in a way that points inference to a local Ollama instance, optionally selecting a model via a command-line flag. In testing, a quantized Qwen 3.5 “35 billion parameter” mixture-of-experts model was used as the local engine, and it successfully handled repository navigation and targeted code analysis.

The most telling results came from how the model performed on different kinds of requests. When asked to understand the repository as a whole, the model struggled to produce a true project-level overview. Instead, it fixated on a single uncommitted file, producing a well-formatted explanation of that file but not the broader “what is my project about?” answer. That mismatch matters because it highlights a limitation of smaller local models: they may excel at local context and formatting, yet fail to synthesize a whole codebase.

When the prompts were directed at a specific file—such as a Python module in an agents directory—the behavior improved sharply. The model read the file, evaluated most of it, and even compared it to a related “trader gate” file. It also surfaced a potentially important issue: a “max steps” or iteration-limit behavior that sounded like a bug. The user then tested whether the model could correct the problem by switching into an auto-edit exception mode (triggered with Shift+Tap). After the edit attempt, the script’s behavior became more explicit: instead of silently returning response content when the agent hit an iteration limit, the agent now reported the limit more clearly. The changes were described as small but meaningful—adding five lines and removing two—suggesting the model could make surgical fixes rather than broad rewrites.

Overall, the local Claude Code + Ollama + Qwen 3.5 workflow looked viable for consumer hardware, at least for tasks that can be grounded in specific files and concrete edits. The presenter’s takeaway was cautious but optimistic: with a 35B quantized mixture-of-experts model, results were “pretty good” for targeted analysis and patching, while repository-wide understanding may require a larger model (for example, 120B-class) to be more reliable. The core message is that local inference can replace hosted Claude Code in many day-to-day coding workflows—provided expectations match the model’s context and synthesis limits.