There is no single hardware requirement for every local model. The practical fit depends on model architecture, parameter count, quantization, context length, runtime, and other work happening on the computer. Use this guide as a way to compare trade-offs, not as a promise that any particular model will run well on every device.
Model size is only one part of the memory requirement. Model weights take memory, and the runtime also needs working space for the context and the current request. Longer conversations, larger context settings and concurrent applications can raise memory use. A quantized model stores weights in a lower-precision representation to reduce memory needs; depending on the model and task, this may affect output quality.
Check the model's published requirements and license, then leave room for the operating system and other programs. Cursiv's model list records the currently offered models and their license notes.
A discrete graphics card is not required for all local inference. The runtime can use the processor, but generation is often slower, especially for larger models. A smaller model and modest context can make CPU use more practical. Keep expectations tied to your exact processor and model.
A GPU can accelerate inference when the runtime can use it, but the model and working data must fit within available graphics memory or be split across GPU and system memory. VRAM demand varies with quantization, context length and runtime. A card's advertised memory is not all available to the model.
System RAM supports the operating system, runtime and any model layers kept outside GPU memory. Disk space is needed for the application, runtime and downloaded model files. The Cursiv README lists 8 GB system RAM and about 6 GB free disk for one model as a starting requirement; additional models need additional storage.
Smaller models generally need fewer resources and can respond faster, but may be less capable on demanding tasks. Larger models can handle some tasks better, while requiring more memory and often taking longer to load or generate. Quantization can lower memory use, but does not make hardware limits disappear.
Long context settings also use resources. If a model is slow or fails to load, try closing other GPU-heavy applications, choosing a smaller compatible model, shortening the context, or using CPU inference with realistic expectations.
Local-model answers can still be wrong. A larger model or faster computer does not guarantee factual accuracy.
Cursiv's hardware notes report one test computer: an Intel i5-12450HX, 32 GB system RAM and an RTX 2050 with 4 GB graphics memory. For one prompt of about 120 words, qwen3:4b-instruct took 3.7 seconds with GPU inference and 13.4 seconds on processor only; qwen2.5:1.5b took 2.6 and 8.3 seconds respectively.
These are single-machine measurements with the model already loaded for the reported response. They do not predict performance on other computers. Read the full measurement notes.
For Cursiv, setup supports local Ollama models, with model and hardware options described in the project's model documentation. Learn more about offline AI in Cursiv, browse the benchmark method, or download Cursiv. For field workflow status, see the project direction.
Cursiv