Lemonade — writing and research https://siplemona.de/blog/ Notes and research from Lemonade, a local AI desktop assistant for Windows and Linux. en Mon, 03 Aug 2026 09:00:00 +0000 Nanite for Experts: a fidelity-first memory hierarchy for local mixture-of-experts inference https://siplemona.de/research/nanite-for-experts https://siplemona.de/research/nanite-for-experts Mon, 03 Aug 2026 09:00:00 +0000 An engineering research report on treating NVMe, RAM, pinned host memory and VRAM as one coordinated hierarchy so a 35B sparse model runs on a 10 GB consumer GPU — with measurements, and the caching ideas that turned out to be slower than getting layer placement right. How much VRAM do you need to run a local LLM? https://siplemona.de/blog/how-much-vram-for-a-local-llm https://siplemona.de/blog/how-much-vram-for-a-local-llm Mon, 03 Aug 2026 09:00:00 +0000 About 0.6 GB per billion parameters at 4-bit, plus the KV cache and roughly a gigabyte you do not get to use — and the throughput cliff we measured when a model nearly fits. Local AI vs cloud AI: what actually changes https://siplemona.de/blog/local-ai-vs-cloud-ai https://siplemona.de/blog/local-ai-vs-cloud-ai Mon, 03 Aug 2026 09:00:00 +0000 Four things genuinely change when the model moves onto your desk, and three things do not change in your favour. What a “local” AI assistant still uses the network for https://siplemona.de/blog/what-local-ai-uses-the-network-for https://siplemona.de/blog/what-local-ai-uses-the-network-for Mon, 03 Aug 2026 09:00:00 +0000 No desktop AI tool sits at zero packets. Every connection Lemonade makes, what it sends, and a two-minute method for checking any vendor's claim yourself.