LM Studio, Ollama suit 8GB rigs

Hot take: this is a model-fit problem first, a tool-choice problem second.

• –The Reddit thread is a privacy-first local assistant question, and the replies quickly shift from app choice to model size and offload math: [reddit thread](https://www.reddit.com/r/LocalLLaMA/comments/1s2c50f/running_llms_with_8_gb_vram_32_gb_ram/).
• –LM Studio's site emphasizes offline local LLMs, local-document chat, and an OpenAI-compatible server, and its CLI exposes GPU-offload controls for tuning VRAM use. In practice, that makes it the friendlier choice if you want to inspect and prompt interactively: [LM Studio](https://lmstudio.ai/), [LM Studio docs](https://lmstudio.ai/docs/cli/local-models/load).
• –Ollama's docs emphasize local execution and a local-only mode, and official model pages still frame 7B as the 8GB RAM floor: [Ollama FAQ](https://docs.ollama.com/faq), [Ollama model page](https://www.ollama.com/library/orca-mini:7b-q5_K_S).
• –Inference from the thread: Qwen3.5-9B, Llama 3.1 8B, and Gemma 2 9B are the practical starts; 30-35B MoE is the stretch goal if you can tolerate slower generations.
• –For scientific questioning, chunking, retrieval, and prompt design will matter more than whether the wrapper is Ollama or LM Studio.