Task Execution Model Selection Discussion for Devices with 32GB RAM and 12GB VRAM

On 2026-11-02, a thread titled "Model selection discussion for small task execution on devices with 32GB memory and 12GB VRAM" appeared on Reddit's LocalLLaMA subreddit, posted by user u/Loose_Doubt367. The poster explained that they want to run two lightweight models on a small local rig to complete tasks, and specified the hardware as 32GB of memory and 12GB of VRAM. Both small models have to fit inside this limited setup, which is the main constraint driving the choice.

The poster laid out a two-tier workflow. Planning is handled by a larger model, either their qwen3.6 35b a3b or a cloud model, while the two small local models do the executing only. These two small models are not supposed to plan at all; they need to be excellent at following instructions and rules precisely. To make room for both small models, the big model is offloaded rather than kept resident, which is why the hardware budget has to be managed carefully.

On the model side, the poster stated they do not want to run strata qwen3.8 flash, and invited other suggestions from the community. This framing treats the setup as a multi-model routing problem: a stronger planner decides what to do, and two small, compliant executors carry out the steps locally. The thread reads as a request for recommendations rather than a report of a settled configuration, and the provided material does not include any follow-up posts, replies, or a final selection. As a result, the discussion currently stands at the point where the poster has described the setup, named the hardware limits, and asked for alternatives to strata qwen3.8 flash, with the outcome still open.