https://geisten.net/research/summary.html Benchmarks, guides and engineering notes on efficient local AI. en https://geisten.net/research/bitnet_pi_reading_speed/https://geisten.net/research/bitnet_pi_reading_speed/Tue, 14 Jul 2026 00:00:00 +0000BitNet b1.58 2B-4T generates 17.4 tokens per second on a Raspberry Pi 5 with the geisten runtime — roughly three times faster than a person reads. The setup, the numbers, and what still goes wrong. https://geisten.net/research/struktur_statt_gpu_reflex/https://geisten.net/research/struktur_statt_gpu_reflex/Sun, 29 Mar 2026 00:00:00 +0000Many AI projects buy compute before they understand their information landscape. A well-structured knowledge map can improve retrieval quality, cost and reliability more than a larger GPU budget. https://geisten.net/research/qwen_small_experts/https://geisten.net/research/qwen_small_experts/Sun, 29 Mar 2026 00:00:00 +0000Three compact Qwen models with clear roles can provide a more useful office workflow than one oversized generalist. The gain comes from task design, local execution and measurable hand-offs. https://geisten.net/research/compliance_deployment_guide/https://geisten.net/research/compliance_deployment_guide/Wed, 25 Mar 2026 00:00:00 +0000Self-hosting an LLM in Europe is only the first step. A dependable system connects licensing, data protection, model provenance, runtime configuration and a signed deployment chain. https://geisten.net/research/ai_future/https://geisten.net/research/ai_future/Sat, 21 Mar 2026 00:00:00 +0000AI is moving from impressive demonstrations into dependable systems. The opportunity is not limited to larger models: it lies in tools, local deployment, specialised workflows and the ability to measure real work. https://geisten.net/research/tool_use_statt_halluzination/https://geisten.net/research/tool_use_statt_halluzination/Wed, 11 Mar 2026 00:00:00 +0000Models become unreliable when they are asked to invent facts that a tool could retrieve exactly. Small models can be effective tool coordinators when the system gives them clear permissions and validated inputs. https://geisten.net/research/rag_small_models/https://geisten.net/research/rag_small_models/Fri, 06 Mar 2026 00:00:00 +0000Retrieval-augmented generation helps models use verified knowledge, but retrieval alone does not guarantee reliable answers. Dependable systems combine source quality, explicit uncertainty and tools for exact facts. https://geisten.net/research/speculative_decoding/https://geisten.net/research/speculative_decoding/Thu, 05 Mar 2026 00:00:00 +0000Speculative decoding pairs a fast draft model with a stronger verifier. The technique can reduce generation latency while preserving the output accepted by the target model. https://geisten.net/research/local_ai_mittelstand/https://geisten.net/research/local_ai_mittelstand/Mon, 02 Mar 2026 00:00:00 +0000For many SMEs, the best AI system is not the largest model but the most dependable architecture: local data boundaries, clear integrations, measurable quality and a runtime that the organisation can operate. https://geisten.net/research/agentic_ux/https://geisten.net/research/agentic_ux/Fri, 13 Feb 2026 00:00:00 +0000A local agent system does not need a swarm of services or a polished dashboard to be useful. This research note outlines a smaller operating model: a terminal interface, explicit tools and one accountable supervisor. https://geisten.net/research/kleine_modelle_grosse_versprechen/https://geisten.net/research/kleine_modelle_grosse_versprechen/Thu, 08 Jan 2026 00:00:00 +0000Model size is not a useful proxy for business value. For narrow tasks, smaller specialists can be faster, less expensive and easier to operate than a single general-purpose model. https://geisten.net/research/germany_ai_actions/https://geisten.net/research/germany_ai_actions/Sat, 27 Dec 2025 00:00:00 +0000Productive AI adoption starts with work, not with a chatbot. These ten actions help SMEs and advisers move from experiments to useful, measurable systems. https://geisten.net/research/parallel_llms/https://geisten.net/research/parallel_llms/Mon, 25 Aug 2025 00:00:00 +0000A single large model is not always the fastest path to a finished result. Measurements on Apple hardware show why independent, small tasks can benefit from parallel execution. https://geisten.net/research/tiny_model_experiment/https://geisten.net/research/tiny_model_experiment/Fri, 15 Aug 2025 00:00:00 +0000A OnePlus 5T from 2017 is enough to demonstrate an important point: modern compact models can run on existing hardware when model format and runtime are chosen carefully. https://geisten.net/research/software_is_everywhere/https://geisten.net/research/software_is_everywhere/Sat, 07 Dec 2024 00:00:00 +0000Layered systems make complex behaviour understandable. The same idea appears in software, natural systems and AI architecture: each layer has a role, an interface and room to adapt. https://geisten.net/research/raspi5_setup/https://geisten.net/research/raspi5_setup/Fri, 06 Dec 2024 00:00:00 +0000A Raspberry Pi 5 is a practical platform for learning what local inference requires. This note uses TinyStories to show how a compact text model can run without a data-centre-class machine. https://geisten.net/research/tinystory/https://geisten.net/research/tinystory/Thu, 05 Dec 2024 00:00:00 +0000TinyStories and Phi-3 show why compact language models matter. They make it possible to study inference, training objectives and deployment constraints without requiring frontier-scale infrastructure. https://geisten.net/research/install_llama3/https://geisten.net/research/install_llama3/Wed, 07 Aug 2024 00:00:00 +0000Ollama provides a simple way to run Llama 3 on a local machine. This short guide covers the minimal setup and the operational questions that matter after the first prompt.