I tested 3 local LLMs on my RTX 4070 Ti for real work — only one earned a permanent spot
This title could be clearer and more informative.Try out Clickbait Shieldfor free (5 uses left this month).
A hands-on comparison of three local LLMs — DeepSeek-R1 8B, Qwen 3.5 9B, and Gemma 4 E4B — run via Ollama on an RTX 4070 Ti with 12GB VRAM. Each model was tested on the same real-world prompts: homelab troubleshooting, article summarization, Python coding, and conversational recommendations. DeepSeek-R1 8B underdelivered despite its reasoning reputation, hallucinating details and missing context. Qwen 3.5 9B was accurate but excessively verbose, often burying useful answers in thousands of tokens. Gemma 4 E4B stood out by using the least VRAM (~7GB), running fastest (~86–90 tok/s), and consistently delivering concise, contextually aware, well-documented responses. Gemma earned the permanent spot on the author's machine based on practical usefulness rather than benchmark scores or community hype.
