Running OpenClaw against a local Ollama model: what to check before you download 20 GB
A short, practical episode: instead of picking a local model by benchmark reputation, the author asks an assistant to reason about his machine and the tool's requirements , then follows the resulting command sequence. The payload is small — one model choice, a handful of switch commands, one live query — but the method is reusable, and it is the part worth writing down. Evidence note: this…
When evaluating whether a local model is compatible with an agent-style tool like OpenClaw, there are two key considerations: the presence of a required marker within the model's capabilities and the model's context length support. A model must explicitly advertise the tool-calling capability using the "qs" tag, which the assistant can verify before downloading a large model.
Context length is also crucial; the author found that a 256K context window is essential for the agent's long tool traces and document handling. After describing his machine specifications to the assistant, he asked for a recommendation. The assistant considered the model's size, the required marker, and context length constraints before suggesting Qwen 3.5 35B A3B Q4 quantisation.
This model combines a mixture-of-experts architecture where roughly 3B parameters are active per token, making it feasible on local hardware despite its 35B total size. The Q4 quantisation strikes a balance between memory usage and computational speed. After confirming the model choice, the author provided the installation commands for OpenClaw, noting that the base installation is covered elsewhere.
He then switched to the recommended model using the provided commands and verified that the switch was successful. Finally, he tested the model with a real-world query—checking today's gold price—though the performance was notably slower than hosted alternatives. The local model was described as "a bit dumber," but still functional for fixed, repeated tasks.
The episode highlights that while local models are less powerful, they can be very useful once a specific task is defined.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
