{"slug": "fun-local-llm-comparisons-with-gemma-granite-and-qwen", "title": "Fun Local LLM Comparisons with Gemma, Granite, and Qwen", "summary": "Ekorbia v0.2 introduces a comparison-chat mode that runs two to three local large language models against the same prompt in parallel. Testing Gemma 4, IBM Granite 4.1, and Qwen 3.5 on a 32 GB M1 Max MacBook Pro revealed that Granite incorrectly selected Chicago deep-dish pizza as a runner-up to New York City's best pizza, while Gemma adopted a \"grumpy librarian\" voice when explaining Hacker News. The feature requires significant memory, as running three large models in parallel can exceed 32 GB and cause swapping that slows performance.", "body_md": "[← All posts](./index.html)\n\n#\nFun local LLM comparisons with\n*Gemma, Granite,* and *Qwen*\n\nEkorbia\n**v0.2**\nfeatures a comparison-chat mode that runs 2-3 local\nmodels against the same prompt in parallel. Here are a\nfew fun prompts running across Gemma 4 (e2b), IBM\nGranite 4.1 (8B), and Qwen 3.5 (4B) on my 32 GB M1 Max\nMacBook Pro.\n\n## 1. The Pizza Question\n\nNew York City is widely regarded as having the best pizza due to its iconic thin-crust style.\n\nThe models were initially reluctant to give a single answer until I attached the following additional prompt: \"Provide clear, concise, opinionated answers to comparison or 'best' questions. Each comparison should have a single winner and a runner-up with a short explanation.\"\n\nThe New York City and Naples answers are acceptable but Granite is clearly wrong here with the runner-up of Chicago deep-dish pizza! And no mention of New Haven style pizza anywhere?\n\n## 2. Explain Hacker News\n\nIt’s a sprawling, perpetually messy digital common room.\n\nGemma is the most fun here, carrying the 'grumpy librarian' voice across multiple paragraphs while Granite and Qwen provide more serious answers with a sprinkling of grumpy librarian at the beginning and the end.\n\n## 3. Will robots take over?\n\nThere is no consensus among experts that unchecked AI growth will inevitably lead to a robot takeover of Earth.\n\nAll three models take the question seriously and none think we are doomed to a Terminator like future.\n\n## Things to watch for with Ekorbia comparison mode.\n\n-\n**Memory matters.** Three large models running in parallel can blow past 32 GB on my MacBook. Ollama will swap them in and out, which makes the \"parallel\" feel serial. -\n**First-token latency varies wildly.** A column that's still showing dots while another is mid-paragraph isn't broken — it's cold-loading. - Granite 4.1 (8B) is fast. It's worth a try if you've mostly been using Qwen or Gemma.\n\n## Send us yours\n\nGot a prompt that produces a hilarious three-way\ndisagreement? Open an\n[issue](https://github.com/ekorbia/ekorbia-desktop/issues)\nwith the prompt and the three outputs and we'll feature\nthe best ones in a follow-up.", "url": "https://wpnews.pro/news/fun-local-llm-comparisons-with-gemma-granite-and-qwen", "canonical_source": "https://ekorbia.com/blog/2026-05-25-fun-local-llm-comparisons", "published_at": "2026-05-28 12:50:02+00:00", "updated_at": "2026-05-28 12:58:39.944501+00:00", "lang": "en", "topics": ["large-language-models", "artificial-intelligence", "generative-ai", "natural-language-processing", "ai-tools"], "entities": ["Gemma", "Granite", "Qwen", "Ekorbia", "IBM", "New York City", "Naples", "Chicago"], "alternates": {"html": "https://wpnews.pro/news/fun-local-llm-comparisons-with-gemma-granite-and-qwen", "markdown": "https://wpnews.pro/news/fun-local-llm-comparisons-with-gemma-granite-and-qwen.md", "text": "https://wpnews.pro/news/fun-local-llm-comparisons-with-gemma-granite-and-qwen.txt", "jsonld": "https://wpnews.pro/news/fun-local-llm-comparisons-with-gemma-granite-and-qwen.jsonld"}}