Local Coding Agents: Qwen3.6, Ollama setup and benchmarks

Sebastian Raschka published a tutorial on Jun 27, 2026 showing how to set up a fully local coding agent using open-weight models. The walkthrough centers on Qwen3.6 35B-A3B served with Ollama and run inside a coding harness such as Qwen-Code, with concrete commands and a simple speed and memory benchmark.

How do you run a local coding agent with Qwen3.6?

Run a locally served model with an inference engine, then connect a coding harness that can read files, run commands, and verify edits. Raschka’s recommended stack for the tutorial uses Ollama to host the LLM and Qwen-Code as the coding client; he gives the exact pull commands and recommends MLX builds on macOS for Apple Silicon.

Install Ollama, then download the model, for example:

• ollama pull qwen3.6:35b-mlx (macOS, MLX build)
• ollama pull qwen3.6:35b (Linux, non-MLX build)

Raschka describes using the Ollama GUI or the command line to confirm the model is running and suggests the /bye command to exit a session. He also notes alternatives to Ollama and Qwen-Code exist, such as LM Studio, vLLM, SGLang, MLX, and other harnesses like Codex and Claude Code, but the tutorial focuses on Ollama plus Qwen-Code for accessibility and speed.

How does Qwen3.6 perform and what resources does it need?

Qwen3.6 35B-A3B is compact enough for local use on capable machines yet needs nontrivial resources: the model is about 22 GB to download and requires roughly 30-40 GB of RAM. Raschka says the 35B-A3B build runs “pretty swiftly” on both a Mac Mini with M4 and a DGX Spark.

For a quick local assessment Raschka provides a speed and memory script (ollama_speed_memory_bench.py) that sends prompts from 1k to 50k words and asks for up to 8k tokens to measure prefill speed, generation speed, and memory use. He gives runnable examples:

• uv run speed-memory-benchmark/ollama_speed_memory_bench.py --model qwen3.6:35b-mlx (macOS)
• uv run speed-memory-benchmark/ollama_speed_memory_bench.py --model qwen3.6:35b (Linux)

Raschka also points to comparative benchmarks: a Polar paper (May 2026) shows the Qwen3.5-4B base model had the best coding performance in the Qwen-Code harness, and Cohere’s June benchmarks show Qwen3.6 35B-A3B "dominates all but one benchmark in this size class." He highlights North Mini Code 1.0 as the most interesting alternative in the same size class.

Why it matters

A fully local coding agent shifts control of data, costs, and reproducibility back to the user. Raschka emphasizes privacy and predictability: local models avoid sending sensitive data to third-party APIs and sidestep API price changes or throttling, citing that Anthropic recently throttled their flagship model’s performance for LLM research. Running models locally also makes workflows reproducible across upgrades, and lets developers modify the harness freely.

What to watch

Watch subsequent benchmark updates and harness-specific tuning: the Polar May 2026 paper and Cohere’s June benchmarks are already influencing which model-harness pairings perform best, and Raschka calls out that Qwen models are often optimized for Qwen-Code. Next signals to monitor are head-to-head speed and stability results from the provided ollama_speed_memory_bench.py runs and any public benchmarks comparing Qwen3.6 35B-A3B and North Mini Code 1.0 in Qwen-Code.